@bendyline/gezel-inference 0.0.0-stage → 1.0.0

package/dist/providers/mlx/python/gezel_mlx_server.py

@@ -0,0 +1,4491 @@
+"""
+gezel_mlx_server — thin OpenAI-compatible HTTP server around mlx-vlm
+that owns the prompt-cache lifecycle.
+
+Why this exists: mlx_vlm.server (the upstream binary) clears the
+prompt cache after every streamed response (see its `finally`:
+`mx.clear_cache()` + `gc.collect()`). That means every turn re-prefills
+the entire context — on a 25K-token Meester session at ~250 tok/s,
+that's a 2-minute wait per turn. This server preserves the cache
+between requests when the caller supplies a `cache_id`, and reuses any
+matching prefix on the next request.
+
+Reuse mechanism: mlx-vlm exposes `PromptCacheState` (in
+mlx_vlm/generate.py) for exactly this purpose — pass it as the
+`prompt_cache_state` kwarg to `stream_generate` and only the new
+tokens beyond the common prefix get prefilled. We hold one
+`PromptCacheState` per `cache_id` in a module-level dict, with bounded
+LRU eviction by total estimated bytes. The batched path plans its own
+reuse via `cache_seed` (longest-common-prefix + protocol trim, plus an
+end-of-prompt KV snapshot for windowed models whose rotated caches
+can't rewind) — see cache_seed.py's module docstring for the why.
+
+Compatibility contract: we serve the same OpenAI-shape
+`/v1/chat/completions` endpoint as mlx_vlm.server, with the same SSE
+shape — the gezel `MlxSession` SSE reader and our existing tool-call
+stripper / Gemma-format parser keep working unchanged. We do NOT do
+server-side tool-call extraction (mlx_vlm.server's `process_tool_calls`
+path); the gezel side already handles that better. The model's
+textual tool-call markup flows through as content; the gezel
+`LeakyToolCallStripper` extracts and repairs it.
+
+Vision requests use the full installed tower, serialized against text batch
+waves. Their KV is deliberately uncached so image identity cannot be lost in
+text-prefix reuse. Ordinary text retains the existing batched cache path.
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import collections
+import hashlib
+import json
+import os
+import re
+import sys
+import threading
+import time
+import traceback
+import types
+import uuid
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+import mlx.core as mx
+from fastapi import FastAPI, HTTPException, Request
+from fastapi.responses import JSONResponse, StreamingResponse
+from pydantic import BaseModel
+
+# mlx_vlm internals we lean on. Importing utils.load + prompt_utils +
+# generate.stream_generate gives us the same model loading, chat
+# templating, and generation primitives as the upstream server. The
+# PromptCacheState class is the single bit of upstream-public API
+# that does the work — see its docstring in mlx_vlm/generate.py.
+from mlx_vlm.generate import PromptCacheState, stream_generate
+from mlx_vlm.prompt_utils import apply_chat_template
+from mlx_vlm.utils import load
+
+# Local module — disk persistence for PromptCacheState.
+# Imported by relative path so the file works whether the server is
+# launched as a module or as a free script (Electron's bundle layout
+# extracts both .py files into the same dir).
+sys.path.insert(0, str(Path(__file__).parent))
+import cache_persist  # noqa: E402
+import think_budget  # noqa: E402
+import cache_seed  # noqa: E402
+import tool_grammar  # noqa: E402
+import spec_decode  # noqa: E402
+import tool_call_stream  # noqa: E402
+import template_stability  # noqa: E402
+import vision_inputs  # noqa: E402
+import wave_policy  # noqa: E402
+from lfm2_compat import ensure_lfm2_config_compat  # noqa: E402
+from qwen3_5_text_compat import is_text_only_qwen3_5_checkpoint  # noqa: E402
+from qwen4_ple import prepare_external_ple_view  # noqa: E402
+
+
+def log_contained_exception(tag: str) -> None:
+    """Print the current traceback for a failure the caller has CONTAINED.
+
+    Tagged line by line, and never `traceback.print_exc()`: the supervisor's
+    stdout classifier reads a bare `SomeError: <message>` leaf line as proof
+    the engine is dead and tears the process down. That is right during
+    startup, which is what it was written for, and wrong at every call site
+    here — each one has already answered the affected request and left the
+    engine serving everyone else on it. A speculative-decode wave that failed
+    one request this way SIGKILLed a 27B three minutes into serving and took
+    an unrelated session's turn down with it.
+    """
+    for line in traceback.format_exc().rstrip().splitlines():
+        print(f"[{tag}] | {line}", flush=True)
+
+
+# ───────── Leaked special-token scrub ─────────
+
+# Chat-template framing tokens — turn / sequence / tool-response
+# delimiters that tight quants (e.g. gemma4-e4b-q4) sometimes emit as
+# literal special-token text. `tokenizer.decode()` keeps special tokens,
+# so these render into the SSE content stream verbatim; wild-caught case
+# was gemma4-e4b-q4 streaming `<eos><|tool_response><eos>` after firing
+# its tool calls. We scrub them at every detokenization site so they
+# never reach the client.
+#
+# Deliberately NOT scrubbed: the tool-CALL markers `<|tool_call>` /
+# `<tool_call|>` (and `<|tool>` / `<tool|>`). The gezel-side
+# LeakyToolCallStripper relies on those flowing through as text to
+# salvage and repair tool calls — this server intentionally does no
+# server-side tool-call extraction (see module docstring). A blanket
+# `skip_special_tokens=True` would delete them too, which is why we
+# strip a targeted denylist instead.
+_LEAKED_MARKER_RE = re.compile(
+    r"<\|?(?:eos|bos|pad|unk|mask|turn|tool_response|start_of_turn|end_of_turn|im_start|im_end)\|?>",
+    re.IGNORECASE,
+)
+
+
+def _make_tool_stream(request_id: str, tools=None):
+    """Build a splitter for this request, or None to keep the old behaviour.
+
+    Returns None whenever the flag is off OR the model's chat template
+    matches no mlx_vlm tool parser — in both cases the caller streams raw
+    content and the gezel-side salvage path stays in charge. Never guess a
+    marker pair: a wrong boundary would silently eat visible output.
+    """
+    if not getattr(ARGS, "stream_tool_calls", False):
+        return None
+    markers = tool_call_stream.resolve_tool_markers(PROCESSOR)
+    if markers is None:
+        return None
+    start, end = markers
+    return tool_call_stream.ToolCallStreamState(
+        start_marker=start, end_marker=end, request_id=request_id, tools=tools
+    )
+
+
+def _tool_stream_fragments(state, chunk_text: str):
+    """Fragments to emit for a chunk; empty when extraction is disabled."""
+    if state is None:
+        return []
+    return state.feed(chunk_text)
+
+
+def _tool_stream_payload(frag, request_id: str, model: str, chunk) -> Dict[str, Any]:
+    """Wrap one fragment in the same chunk envelope the content path uses."""
+    if frag.content is not None:
+        delta: Dict[str, Any] = {"role": "assistant", "content": frag.content}
+    else:
+        call: Dict[str, Any] = {
+            "index": frag.tool_call_index,
+            "function": {"arguments": frag.arguments or ""},
+        }
+        if frag.tool_call_id is not None:
+            call["id"] = frag.tool_call_id
+            call["type"] = "function"
+            # The grammar fixes the name inside the argument body, so the
+            # name field stays empty rather than being guessed here; the
+            # client's parser reads it from the arguments as it always has.
+            call["function"]["name"] = frag.name or ""
+        delta = {"role": "assistant", "tool_calls": [call]}
+    return {
+        "id": request_id,
+        "object": "chat.completion.chunk",
+        "created": int(time.time()),
+        "model": model,
+        "choices": [{"index": 0, "delta": delta, "finish_reason": None}],
+        "usage": {
+            "input_tokens": getattr(chunk, "prompt_tokens", 0),
+            "output_tokens": getattr(chunk, "generation_tokens", 0),
+            "total_tokens": (
+                getattr(chunk, "prompt_tokens", 0) + getattr(chunk, "generation_tokens", 0)
+            ),
+            "prompt_tps": getattr(chunk, "prompt_tps", 0.0),
+            "generation_tps": getattr(chunk, "generation_tps", 0.0),
+            "cached_tokens": getattr(chunk, "cached_tokens", 0),
+        },
+    }
+
+
+def _scrub_leaked_markers(text: str) -> str:
+    """Drop chat-template framing special-tokens that leaked as text."""
+    if not text or "<" not in text:
+        return text
+    return _LEAKED_MARKER_RE.sub("", text)
+
+
+def _generation_config_eos_ids(model_dir: str) -> "List[int]":
+    """Read the stop-token set a model declares in generation_config.json.
+
+    mlx's tokenizer wrapper doesn't always surface the full eos set the
+    model author declared. gemma4-e4b-q4 is the wild-caught case: it
+    declares `eos_token_id: [1, 106, 50]` (`<eos>`, `<turn|>`,
+    `<|tool_response>`), but the wrapper resolved a narrower set, so the
+    batch engine never stopped on `<eos>` / `<|tool_response>` and
+    streamed them as text after the model's tool calls. Merging these in
+    halts generation at the real turn boundary instead of running on
+    into framing tokens (the upstream fix that complements the scrub).
+    """
+    try:
+        with open(os.path.join(model_dir, "generation_config.json")) as f:
+            cfg = json.load(f)
+    except (OSError, ValueError):
+        return []
+    eos = cfg.get("eos_token_id")
+    if isinstance(eos, bool):  # guard: bools are ints in Python
+        return []
+    if isinstance(eos, int):
+        return [eos]
+    if isinstance(eos, list):
+        return [int(e) for e in eos if isinstance(e, int) and not isinstance(e, bool)]
+    return []
+
+
+def _tool_call_close_stop_ids(model_dir: str) -> "List[int]":
+    """`</tool_call>` special-token ids to halt a turn at the first closed
+    tool call — for models that expect a server-side tool-call parser we
+    don't run.
+
+    Poolside Laguna (and kin) declare `tool_call_parser` in
+    generation_config.json and emit `<tool_call>…</tool_call>` WITHOUT
+    listing `</tool_call>` in `eos_token_id`: their production server (the
+    `*_v1` parser) detects the closed call and stops generation. On
+    mlx_vlm's raw completion path nothing stops it, so the model runs on
+    past the first call and loops out hundreds of calls in a single turn
+    until the SSE stalls (wild-caught: laguna-s-118b/MLX — ~450
+    `<tool_call>` spans in one turn, then a 300s silent stall → the whole
+    turn aborts and no call lands). Making the `</tool_call>` special token
+    a stop halts the turn at the first closed call; the gezel GLM salvage
+    layer promotes it and the agentic loop advances one call per turn.
+
+    Gated on the model DECLARING a non-empty `tool_call_parser` so models
+    that end tool-call turns on their own eos (Qwen/Gemma) are untouched —
+    zero regression risk for the already-tuned cast.
+    """
+    try:
+        with open(os.path.join(model_dir, "generation_config.json")) as f:
+            cfg = json.load(f)
+    except (OSError, ValueError):
+        return []
+    parser = cfg.get("tool_call_parser")
+    if not isinstance(parser, str) or not parser:
+        return []
+    try:
+        with open(os.path.join(model_dir, "tokenizer_config.json")) as f:
+            tcfg = json.load(f)
+    except (OSError, ValueError):
+        return []
+    added = tcfg.get("added_tokens_decoder") or {}
+    ids: "List[int]" = []
+    for sid, meta in added.items():
+        if isinstance(meta, dict) and meta.get("content") == "</tool_call>":
+            try:
+                ids.append(int(sid))
+            except (TypeError, ValueError):
+                pass
+    return ids
+
+
+# ───────── Configuration (CLI args) ─────────
+
+parser = argparse.ArgumentParser(description="gezel MLX wrapper server")
+parser.add_argument("--model", required=True, help="Path to the MLX model directory")
+parser.add_argument(
+    "--external-ple-dir",
+    type=str,
+    default=None,
+    help=(
+        "Derived model-view directory for Qwen4's disk-backed PLE table. "
+        "Applicable checkpoints are hard-linked into this view and the large "
+        "n-gram table is memory-mapped by mlx-vlm 0.7.1. Other model families "
+        "keep the source directory; an eligible Qwen4 checkpoint fails safely "
+        "when the configured runtime cannot prepare the view."
+    ),
+)
+parser.add_argument(
+    "--allow-resident-ple-fallback",
+    action="store_true",
+    help=(
+        "Advanced/manual override: if a quantized Qwen4 checkpoint is eligible "
+        "for --external-ple-dir but the derived view cannot be prepared, load "
+        "the full PLE table resident. Gezel never passes this because its "
+        "catalog working-set reservation assumes external PLE is active."
+    ),
+)
+parser.add_argument("--host", default="127.0.0.1")
+parser.add_argument("--port", type=int, default=8080)
+parser.add_argument(
+    "--cache-budget-mb",
+    type=int,
+    default=4096,
+    help="Bound on total in-memory prompt-cache size before LRU eviction.",
+)
+parser.add_argument(
+    "--prefill-step-size",
+    type=int,
+    default=2048,
+    help="Tokens per prefill step. Lower = less peak memory, slower.",
+)
+parser.add_argument(
+    "--spec-draft-model",
+    type=str,
+    default=None,
+    help="Path to an MTP drafter directory; enables speculative decoding "
+    "for eligible (greedy, processor-free) requests. See spec_decode.py.",
+)
+parser.add_argument(
+    "--spec-draft-kind",
+    type=str,
+    default=None,
+    help="Drafter kind override (default: auto-detect from the drafter's model_type).",
+)
+parser.add_argument(
+    "--spec-block-size",
+    type=int,
+    default=None,
+    help="Draft block size override (default: the drafter's configured block_size).",
+)
+parser.add_argument(
+    "--persist-dir",
+    type=str,
+    default=None,
+    help=(
+        "Directory for on-disk PromptCacheState persistence. When set, "
+        "in-memory cache misses fall back to a disk lookup, evictions "
+        "save before dropping, and graceful shutdown flushes all warm "
+        "entries. Disabled when unset (matches Phase 2 behavior)."
+    ),
+)
+parser.add_argument(
+    "--model-fingerprint",
+    type=str,
+    default=None,
+    help=(
+        "Stable identifier for the loaded model — sha256-derived from the "
+        "manifest. Disk-cache entries are segmented by fingerprint so a "
+        "model swap can never load wrong-shape KV state. Required when "
+        "--persist-dir is set."
+    ),
+)
+parser.add_argument(
+    "--disk-cache-budget-mb",
+    type=int,
+    default=8192,
+    help=(
+        "Bound on total on-disk persisted cache size. LRU pruning runs "
+        "after every save when over budget. 0 disables pruning."
+    ),
+)
+parser.add_argument("--vision", action="store_true", help="Load and enable the installed vision tower")
+parser.add_argument(
+    "--max-concurrency",
+    type=int,
+    default=1,
+    help=(
+        "Max sequences to generate concurrently. Any positive value routes "
+        "generation through a single mlx_lm BatchGenerator; 1 (default) is a "
+        "singleton queue that still uses the cache-safe prompt-boundary "
+        "snapshot path, while N>1 also enables continuous batching. Set 0 only "
+        "for the legacy serial stream_generate fallback. Set from the TS "
+        "provider's batchMaxConcurrency."
+    ),
+)
+parser.add_argument(
+    "--gpu-memory-limit-mb",
+    type=int,
+    default=0,
+    help=(
+        "Total GPU/unified-memory budget (MB) this engine may use — weights + "
+        "KV + compute. Set from the TS capacity broker (min of the model budget "
+        "and Metal's recommended working set). Drives memory-aware wave "
+        "admission: when live usage nears this ceiling the BatchGenerator "
+        "collapses toward serial instead of co-batching concurrent prefills "
+        "into a Metal command-buffer OOM (which SIGABRTs the whole process). "
+        "0 (default) = unset → admit purely on --max-concurrency, as before."
+    ),
+)
+parser.add_argument(
+    "--wired-limit-mb",
+    type=int,
+    default=0,
+    help=(
+        "How much this engine may PIN (MB) — the planned resident set: weights "
+        "plus the KV its slots will hold. Deliberately NOT the same number as "
+        "--gpu-memory-limit-mb: that is a ceiling on what may be USED, while "
+        "wired pages are unreclaimable by the OS, so wiring the whole share "
+        "starves the machine (a 27B wired 96 GiB for a ~45 GB working set and "
+        "left 155 MB free on a 128 GB Mac). 0 (default) = leave the wired "
+        "limit alone."
+    ),
+)
+parser.add_argument(
+    "--stream-tool-calls",
+    action="store_true",
+    help=(
+        "Emit OpenAI-shaped `tool_calls` argument deltas instead of streaming "
+        "the model's tool markup as content. OFF by default: the gezel-side "
+        "salvage path stays the fallback for models whose chat template "
+        "matches no mlx_vlm tool parser. On, the client gets a real tool-call "
+        "boundary and can end a turn once the call is usable — the same lever "
+        "llama-server gives us, and the reason MLX turns otherwise run to "
+        "completion (measured 41 turns at a 136s median, 203 min/10 trials)."
+    ),
+)
+parser.add_argument(
+    "--kv-bits",
+    type=int,
+    default=0,
+    help=(
+        "Quantize the KV cache to this many bits. 0 (default) = f16. Halves "
+        "(8-bit) or quarters (4-bit) the largest term in a long-context "
+        "session: a 256K window on a 16-full-attention-layer 27B is 16 GiB of "
+        "f16 KV. Applies to the SERIAL generation path only — mlx-lm's "
+        "BatchGenerator builds BatchKVCache layers, which have no "
+        "`to_quantized`, and `maybe_quantize_kv_cache` skips them silently. "
+        "The launcher therefore plans memory at f16 whenever batching is on, "
+        "so the broker never reserves a discount the engine did not take."
+    ),
+)
+parser.add_argument(
+    "--kv-group-size",
+    type=int,
+    default=64,
+    help="Group size for KV quantization. mlx-lm's default is 64.",
+)
+parser.add_argument(
+    "--kv-quant-scheme",
+    default="uniform",
+    help=(
+        "Quantization scheme passed to mlx-vlm. 'uniform' is right for integer "
+        "bit values; mlx-vlm selects TurboQuant for fractional ones."
+    ),
+)
+parser.add_argument(
+    "--quantized-kv-start",
+    type=int,
+    default=0,
+    help=(
+        "Only quantize a layer once its cache passes this many tokens. Leaves "
+        "short turns on the f16 path, where quantization costs more than the "
+        "memory it saves."
+    ),
+)
+ARGS = parser.parse_args()
+
+
+# Resolve the persistence directory once. None means disk persistence
+# is disabled — every code path that touches disk gates on this.
+PERSIST_DIR: Optional[Path] = (
+    Path(ARGS.persist_dir).expanduser() if ARGS.persist_dir else None
+)
+MODEL_FINGERPRINT: Optional[str] = ARGS.model_fingerprint
+if PERSIST_DIR is not None and not MODEL_FINGERPRINT:
+    print(
+        "[cache] --persist-dir set without --model-fingerprint; refusing to "
+        "persist (mismatched fingerprints would silently corrupt outputs).",
+        flush=True,
+    )
+    PERSIST_DIR = None
+
+
+def _disk_enabled() -> bool:
+    return PERSIST_DIR is not None and MODEL_FINGERPRINT is not None
+
+
+# ───────── Parent-death watchdog ─────────
+
+
+def _start_parent_death_watchdog() -> None:
+    """Self-terminate if our parent process goes away.
+
+    Why: this server holds 10–20 GB of model weights in unified memory.
+    If the gezel daemon crashes, is force-quit, or is SIGKILL'd by
+    VSCode's debug "Stop" button, we get reparented to launchd (pid 1)
+    and sit on that RAM forever — orphan instances accumulate every
+    dev iteration.
+
+    macOS has no `prctl(PR_SET_PDEATHSIG)` equivalent; polling
+    `getppid()` from a daemon thread is the portable substitute. The
+    GIL is released during MLX/torch native compute and during uvicorn
+    epoll waits, so the 2s tick still fires reliably even under load.
+    A `daemon=True` thread doesn't block interpreter shutdown when the
+    main thread exits cleanly — the watchdog only matters for the
+    crash path.
+    """
+    initial_ppid = os.getppid()
+    if initial_ppid == 1:
+        # Already orphaned at startup — there's no parent to serve.
+        # Refuse to load weights only to immediately leak them.
+        print(
+            "[gezel-mlx] parent already pid=1 at startup; refusing to load model",
+            flush=True,
+        )
+        os._exit(1)
+
+    def watch() -> None:
+        while True:
+            time.sleep(2.0)
+            current = os.getppid()
+            if current == initial_ppid:
+                continue
+            # Reparented — original parent is gone. The new ppid is
+            # almost always 1 (launchd) or a session leader; either way
+            # nobody who started us is still listening, so release the
+            # model RAM. `os._exit` skips atexit/finalizers — uvicorn's
+            # graceful-shutdown path would try to drain in-flight
+            # streams against a dead client and stall, which defeats
+            # the point of the watchdog.
+            print(
+                f"[gezel-mlx] parent vanished (was {initial_ppid}, now {current}); exiting",
+                flush=True,
+            )
+            os._exit(0)
+
+    t = threading.Thread(
+        target=watch, name="gezel-parent-death-watchdog", daemon=True
+    )
+    t.start()
+
+
+_start_parent_death_watchdog()
+
+
+# ───────── Model loading (one-shot at startup) ─────────
+
+
+print(f"Started server process [{os.getpid()}]", flush=True)
+print("Waiting for application startup.", flush=True)
+_SOURCE_MODEL_DIR = ARGS.model
+ARGS.model, _ple_status, _ple_applicable = prepare_external_ple_view(
+    ARGS.model, ARGS.external_ple_dir
+)
+if ARGS.model != _SOURCE_MODEL_DIR:
+    print(f"[ple] external view {_ple_status}: {ARGS.model}", flush=True)
+elif ARGS.external_ple_dir:
+    if not _ple_applicable:
+        print(f"[ple] not applicable: {_ple_status}", flush=True)
+    elif ARGS.allow_resident_ple_fallback:
+        print(f"[ple] resident fallback explicitly allowed: {_ple_status}", flush=True)
+    else:
+        raise RuntimeError(
+            "external Qwen4 PLE view is required by the planned memory footprint "
+            f"but could not be prepared: {_ple_status}"
+        )
+print(f"Loading model from: {ARGS.model}", flush=True)
+_LFM2_FF_DIM = ensure_lfm2_config_compat(ARGS.model)
+if _LFM2_FF_DIM is not None:
+    print(f"Applied lfm2 config compat: block_ff_dim={_LFM2_FF_DIM}", flush=True)
+_QWEN3_5_TEXT_ONLY = is_text_only_qwen3_5_checkpoint(ARGS.model)
+if _QWEN3_5_TEXT_ONLY:
+    print(
+        "Detected text-only Qwen 3.5 checkpoint; allowing absent vision-tower weights.",
+        flush=True,
+    )
+
+# ── Tower selection (2026-08-28) ──
+#
+# Text-only launches can use the faster language tower. --vision retains
+# the full VLM tower for pixel encoding and serialized multimodal requests.
+# Yet mlx_vlm's language towers pay a large context-growing decode tax
+# under the BatchGenerator this server runs on. Measured on
+# qwen3.8-27b-q4 (matched-thermal 2x2, ms per decoded token at
+# 10.9k/54.3k prompt tokens):
+#
+#     mlx_lm tower  x BatchGenerator : 32.1 / 41.2   (= bare serial)
+#     mlx_vlm tower x BatchGenerator : 44.7 / 86.7   (slope 0.97 vs 0.34 us/tok)
+#
+# The vlm tower runs 3-axis multimodal RoPE with per-layer array-offset
+# position ids + left-padding mask machinery per token; mlx_lm's text
+# tower is plain rope(offset) + fused SDPA. Same checkpoint, ~2x decode
+# at working context — so batch mode loads the mlx_lm tower when it can.
+#
+# Scope guards, deliberately conservative:
+#   - allow-listed architectures only (verified by the 2x2 above); other
+#     archs keep the vlm loader until they get their own measurement,
+#   - batch mode only (--max-concurrency >= 1, the default): the legacy
+#     serial path is mlx_vlm's stream_generate and keeps the vlm loader,
+#   - GEZEL_MLX_TEXT_TOWER=off restores the old behavior outright;
+#     =force skips the allowlist (measurement runs).
+# Every branch prints a `[tower]` fingerprint line — an A/B that cannot
+# prove which tower served it is how this class of defect stayed hidden.
+_TEXT_TOWER = None
+_TEXT_TOWER_ALLOWED_ARCHS = {"qwen3_5"}
+
+
+def _model_architecture(model_dir: str) -> str:
+    try:
+        with open(os.path.join(model_dir, "config.json"), "r") as fh:
+            return str(json.load(fh).get("model_type", ""))
+    except Exception:
+        return ""
+
+
+_tower_pref = os.environ.get("GEZEL_MLX_TEXT_TOWER", "auto").strip().lower()
+_arch = _model_architecture(ARGS.model)
+if (
+    not ARGS.vision
+    and _tower_pref not in ("0", "off", "false")
+    and int(getattr(ARGS, "max_concurrency", 1) or 0) >= 1
+    and (_tower_pref == "force" or _arch in _TEXT_TOWER_ALLOWED_ARCHS)
+    # A configured MTP drafter needs the vlm tower (rollback lives there);
+    # an explicit GEZEL_MLX_TEXT_TOWER=force still wins and spec turns off.
+    and not (getattr(ARGS, "spec_draft_model", None) and _tower_pref != "force")
+):
+    try:
+        from mlx_lm.utils import load as _mlx_lm_load
+
+        MODEL, PROCESSOR = _mlx_lm_load(ARGS.model)
+        _TEXT_TOWER = "mlx_lm"
+        print(
+            f"[tower] active=mlx_lm arch={_arch or 'unknown'} "
+            f"(text tower; vlm loader bypassed for batch serving)",
+            flush=True,
+        )
+    except Exception as exc:
+        print(
+            f"[tower] mlx_lm load failed for arch={_arch or 'unknown'}: {exc} "
+            f"— falling back to mlx_vlm",
+            flush=True,
+        )
+if _TEXT_TOWER is None:
+    MODEL, PROCESSOR = load(ARGS.model, strict=not _QWEN3_5_TEXT_ONLY)
+    _reason = (
+        "native vision enabled" if ARGS.vision else
+        "GEZEL_MLX_TEXT_TOWER=off"
+        if _tower_pref in ("0", "off", "false")
+        else "serial mode (--max-concurrency 0)"
+        if int(getattr(ARGS, "max_concurrency", 1) or 0) < 1
+        else "spec drafter configured (MTP runs on the vlm tower)"
+        if (getattr(ARGS, "spec_draft_model", None) and _tower_pref != "force")
+        else f"arch {_arch or 'unknown'} not in text-tower allowlist"
+        if _tower_pref != "force"
+        else "mlx_lm load failed"
+    )
+    print(f"[tower] active=mlx_vlm ({_reason})", flush=True)
+_VISION_ENABLED = bool(ARGS.vision and getattr(PROCESSOR, "image_processor", None) is not None)
+if ARGS.vision and not _VISION_ENABLED:
+    raise RuntimeError("Native vision was requested but this checkpoint has no image processor")
+print(f"[vision] enabled={_VISION_ENABLED}", flush=True)
+print("Model and processor loaded successfully.", flush=True)
+
+# ── MTP speculative decoding (2026-08-28) ──
+# Armed only when a drafter is configured AND the capability probe passes;
+# every failure logs one `[spec] off` line and the server serves normally.
+_SPEC = spec_decode.resolve_spec(ARGS, MODEL, _TEXT_TOWER)
+if _SPEC is not None:
+    print(
+        f"[spec] active drafter={_SPEC.drafter_dir} kind={_SPEC.kind} "
+        f"block={_SPEC.block_size or 'config-default'} tower=mlx_vlm",
+        flush=True,
+    )
+
+
+# Prefix-stability work (template relocation + reasoning-stable snapshot
+# boundaries), OFF by default. Wired, unit-tested, and demonstrably firing —
+# a single turn reused 6,980 tokens against 165 prefilled — but the paired
+# arms measured it as a REGRESSION on its own metric:
+#
+#   reuse%   34 (off) -> 26 -> 22 -> 14 (all fixes on)
+#   prefill/trial 37,486 (off) -> 43,586 (on)
+#
+# Truncating saves at the stable boundary buys matches but shrinks what each
+# match is worth, and warming with tools raises the prefill denominator. Net
+# negative at n=1-5 with ~3x trial spread, so it is neither shippable nor
+# disproven. Same posture as GEZEL_LAYERED_PREFIX_CACHE: keep it measurable.
+_STABLE_PREFIX_ENABLED = os.environ.get("GEZEL_MLX_STABLE_PREFIX") in ("1", "true")
+
+
+def _resolve_stable_chat_template() -> Optional[str]:
+    """Relocate the reasoning preamble below the cache-stable content.
+
+    Qwen 3.5+ templates put the reasoning-effort line ABOVE the tool block,
+    so flipping reasoning mode between turns — which the constrained-write
+    path does deliberately — changed the prompt from token 3 onward and
+    re-prefilled the entire tool block every time. Measured on
+    qwen3.8-27b-q4: 3 shared tokens before, 2,921 after.
+
+    Verified before adoption and abandoned on any doubt: a mis-rendered
+    prompt is far worse than a cold cache.
+    """
+    if not _STABLE_PREFIX_ENABLED:
+        return None
+    tok = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+    original = getattr(tok, "chat_template", None)
+    if not isinstance(original, str) or not original:
+        return None
+    patched = template_stability.relocate_reasoning_preamble(original)
+    if patched is None:
+        print("[template] preamble relocation not applicable to this template", flush=True)
+        return None
+    probe_messages = [
+        {"role": "system", "content": "Your role is Builder.\n\n" + ("about text. " * 64)},
+        {"role": "user", "content": "hi"},
+    ]
+    probe_tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "write_file",
+                "description": "Write a file to the workspace.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"path": {"type": "string"}, "content": {"type": "string"}},
+                },
+            },
+        }
+    ]
+
+    def _render(template: str, kwargs: Dict[str, Any]) -> str:
+        return tok.apply_chat_template(
+            probe_messages,
+            tools=probe_tools,
+            add_generation_prompt=True,
+            tokenize=False,
+            chat_template=template,
+            **kwargs,
+        )
+
+    try:
+        ok, reason = template_stability.verify_prefix_stability(
+            _render, tok.encode, patched, original
+        )
+    except Exception as exc:  # noqa: BLE001 — verification must never block boot
+        print(f"[template] preamble relocation verification errored: {exc}", flush=True)
+        return None
+    if not ok:
+        print(f"[template] preamble relocation REJECTED: {reason}", flush=True)
+        return None
+    print(f"[template] preamble relocated for prefix stability — {reason}", flush=True)
+    return patched
+
+
+STABLE_CHAT_TEMPLATE = _resolve_stable_chat_template()
+
+
+# ───────── GPU memory ceiling (Metal) ─────────
+# Effective ceiling on total unified-memory this engine may use — weights + KV
+# + compute. The batch admission path (BatchEngine._admit_count) reads live
+# usage against it to decide how many concurrent prefills to start, so a wave
+# never piles KV past the point where a Metal command-buffer allocation aborts
+# the whole python process. 0 = unset (no ceiling → admit on width alone).
+_MEM_LIMIT_BYTES = 0
+# The resident-set target supplied by the TS capacity broker. Keep this
+# separately from the allocation ceiling: mlx_lm's BatchGenerator currently
+# overwrites the process wired limit with Metal's full recommended working set
+# during construction, so the Gezel policy must be reapplied afterwards.
+_WIRED_LIMIT_BYTES = 0
+# MLX keeps freed Metal buffers for reuse in a process-global cache. That is
+# useful, but an unbounded pool made a ~62 GB working set retain another
+# ~45 GB until mlx_lm's coarse 512-step clear, producing a 65 -> 120 GB
+# sawtooth. This cap is intentionally distinct from the prompt/KV cache budget.
+_MLX_BUFFER_CACHE_LIMIT_BYTES = 0
+# Throttle a multi-sub wave down to serial once live allocation reaches this
+# fraction of the ceiling. Leaves headroom for the wave's own prefill/compute
+# activation buffers, which land on top of already-resident weights + KV.
+_MEM_ADMIT_FRACTION = 0.85
+
+
+def _safe_active_memory() -> int:
+    """Live MLX allocation in bytes (weights + resident KV), or 0 if the
+    runtime doesn't expose it. Read-only; never raises."""
+    try:
+        return int(mx.get_active_memory())
+    except Exception:  # noqa: BLE001 — introspection is best-effort
+        return 0
+
+
+def _metal_recommended_working_set() -> int:
+    """Metal's `max_recommended_working_set_size` (bytes), or 0 if unavailable.
+    This is the OS's own soft ceiling for GPU-resident memory."""
+    try:
+        # mlx 0.32.2 (resolved by mlx-vlm 0.7.1) deprecates the old
+        # mx.metal.device_info() namespace. Keep the fallback so an operator's
+        # older, explicitly configured mlx package still gets a useful cap.
+        device_info = getattr(mx, "device_info", None)
+        info = device_info() if callable(device_info) else mx.metal.device_info()
+        return int(info.get("max_recommended_working_set_size", 0) or 0)
+    except Exception:  # noqa: BLE001 — API shape varies across mlx versions
+        return 0
+
+
+def _safe_cache_memory() -> int:
+    """MLX's retained free-buffer pool in bytes, or 0 if unavailable.
+
+    Distinct from active memory and invisible to it: MLX holds freed buffers
+    for reuse rather than returning them, so a process can sit far above its
+    live allocation without `get_active_memory` moving. That gap is most of
+    the difference between a model's working set and its physical footprint.
+    """
+    try:
+        return int(mx.get_cache_memory())
+    except Exception:  # noqa: BLE001 — introspection is best-effort
+        return 0
+
+
+def _safe_peak_memory() -> int:
+    """Peak MLX allocation in bytes, or 0 if unavailable."""
+    try:
+        return int(mx.get_peak_memory())
+    except Exception:  # noqa: BLE001 — introspection is best-effort
+        return 0
+
+
+def _log_memory(phase: str, cache_before: Optional[int] = None) -> None:
+    """One line of memory truth per turn.
+
+    Nothing observed this engine between startup and the next restart, so a
+    footprint drifting toward the wired ceiling was only ever reconstructible
+    afterwards from `vm_stat`. `cache` is the number that made the difference
+    invisible: it does not move `active`, and no other layer measures it —
+    the in-engine prompt-cache budget tracks its own token-count estimate.
+    """
+    active = _safe_active_memory()
+    cached = _safe_cache_memory()
+    peak = _safe_peak_memory()
+    if active <= 0 and cached <= 0:
+        return
+    mb = 1024 * 1024
+    ceiling = f", ceiling={_MEM_LIMIT_BYTES // mb}MB" if _MEM_LIMIT_BYTES > 0 else ""
+    wired = f", wired_limit={_WIRED_LIMIT_BYTES // mb}MB" if _WIRED_LIMIT_BYTES > 0 else ""
+    cache_limit = (
+        f", cache_limit={_MLX_BUFFER_CACHE_LIMIT_BYTES // mb}MB"
+        if _MLX_BUFFER_CACHE_LIMIT_BYTES > 0
+        else ""
+    )
+    peak_text = f" peak={peak // mb}MB" if peak > 0 else ""
+    reclaimed_text = (
+        f" cache_before={cache_before // mb}MB"
+        if cache_before is not None and cache_before > cached
+        else ""
+    )
+    print(
+        f"[mem] {phase}: active={active // mb}MB cache={cached // mb}MB "
+        f"total={(active + cached) // mb}MB{peak_text}{reclaimed_text}"
+        f"{ceiling}{wired}{cache_limit}",
+        flush=True,
+    )
+
+
+def _apply_mlx_memory_policy(phase: str) -> None:
+    """Apply Gezel's process-wide MLX memory policy.
+
+    This is deliberately idempotent. In particular it is called immediately
+    after BatchGenerator construction because mlx_lm resets the wired limit in
+    its constructor. Every MLX call is guarded for older sidecar runtimes.
+    """
+    if _WIRED_LIMIT_BYTES > 0:
+        set_wired = getattr(mx, "set_wired_limit", None)
+        if set_wired is not None:
+            try:
+                set_wired(_WIRED_LIMIT_BYTES)
+            except Exception as exc:  # noqa: BLE001
+                print(f"[mem] {phase} set_wired_limit skipped: {exc}", flush=True)
+    if _MLX_BUFFER_CACHE_LIMIT_BYTES > 0:
+        set_cache = getattr(mx, "set_cache_limit", None)
+        if set_cache is not None:
+            try:
+                set_cache(_MLX_BUFFER_CACHE_LIMIT_BYTES)
+            except Exception as exc:  # noqa: BLE001
+                print(f"[mem] {phase} set_cache_limit skipped: {exc}", flush=True)
+
+
+def _reclaim_mlx_buffer_cache(
+    phase: str, force: bool = False, log: bool = False
+) -> bool:
+    """Release MLX's free-buffer pool at a safe boundary or under pressure.
+
+    Active allocations (model weights and live KV state) are untouched.
+    `force` is used once a batch turn is fully drained; during generation we
+    clear only when total MLX memory crosses the admission threshold or the
+    runtime has temporarily exceeded the configured free-buffer cap.
+    """
+    active = _safe_active_memory()
+    cached = _safe_cache_memory()
+    under_pressure = (
+        _MEM_LIMIT_BYTES > 0
+        and active + cached >= _MEM_LIMIT_BYTES * _MEM_ADMIT_FRACTION
+    )
+    over_cache_limit = (
+        _MLX_BUFFER_CACHE_LIMIT_BYTES > 0
+        and cached > _MLX_BUFFER_CACHE_LIMIT_BYTES
+    )
+    reclaimed = False
+    if force or under_pressure or over_cache_limit:
+        clear_cache = getattr(mx, "clear_cache", None)
+        if clear_cache is not None:
+            try:
+                clear_cache()
+                reclaimed = True
+            except Exception as exc:  # noqa: BLE001
+                print(f"[mem] {phase} clear_cache skipped: {exc}", flush=True)
+    # Turn boundaries and actual pressure events are useful telemetry. A
+    # healthy periodic probe should stay silent instead of writing one line
+    # every 32 decode steps.
+    if log or force or under_pressure or over_cache_limit:
+        _log_memory(phase, cache_before=cached if reclaimed else None)
+    return reclaimed
+
+
+def _setup_gpu_memory_ceiling() -> None:
+    """Resolve the effective memory ceiling and the wired limit.
+
+    The ceiling is the smaller of what the supervisor budgeted for this model
+    (`--gpu-memory-limit-mb`, from the TS capacity broker) and what Metal
+    recommends as a working set — whichever is tighter is the honest bound.
+    Everything here is guarded: a version whose mlx lacks these calls must
+    degrade to "no ceiling" (width-only admission), never crash startup.
+
+    The WIRED limit is a separate, smaller number (`--wired-limit-mb`): the
+    planned resident set. Wiring the whole ceiling is what let a 27B hold
+    ~103 GB for a ~45 GB working set — wired pages never come back under
+    pressure, so this must track what steady-state inference actually needs
+    resident, not what the engine is allowed to allocate. Falls back to the
+    ceiling only when the supervisor passed no plan (unreadable geometry).
+    """
+    global _MEM_LIMIT_BYTES, _WIRED_LIMIT_BYTES, _MLX_BUFFER_CACHE_LIMIT_BYTES
+    recommended = _metal_recommended_working_set()
+    arg_limit = max(0, int(getattr(ARGS, "gpu_memory_limit_mb", 0) or 0)) * 1024 * 1024
+    candidates = [b for b in (recommended, arg_limit) if b > 0]
+    _MEM_LIMIT_BYTES = min(candidates) if candidates else 0
+    planned = max(0, int(getattr(ARGS, "wired_limit_mb", 0) or 0)) * 1024 * 1024
+    _WIRED_LIMIT_BYTES = (
+        min(planned, _MEM_LIMIT_BYTES)
+        if planned > 0 and _MEM_LIMIT_BYTES > 0
+        else planned or _MEM_LIMIT_BYTES
+    )
+    # Retain a small amount of free Metal storage for fast reuse, scaled to
+    # the planned working set but never allowed to become a second model-sized
+    # pool. Current 27B geometry resolves to just under 4 GB.
+    cache_scale = _WIRED_LIMIT_BYTES or _MEM_LIMIT_BYTES
+    if cache_scale > 0:
+        _MLX_BUFFER_CACHE_LIMIT_BYTES = max(
+            256 * 1024 * 1024,
+            min(4 * 1024 * 1024 * 1024, cache_scale // 16),
+        )
+    else:
+        _MLX_BUFFER_CACHE_LIMIT_BYTES = 1024 * 1024 * 1024
+    _apply_mlx_memory_policy("startup")
+    print(
+        f"[mem] gpu ceiling={_MEM_LIMIT_BYTES // (1024 * 1024)}MB "
+        f"(budget={arg_limit // (1024 * 1024)}MB, "
+        f"metal_recommended={recommended // (1024 * 1024)}MB, "
+        f"wired={_WIRED_LIMIT_BYTES // (1024 * 1024)}MB, "
+        f"buffer_cache_limit={_MLX_BUFFER_CACHE_LIMIT_BYTES // (1024 * 1024)}MB, "
+        f"active={_safe_active_memory() // (1024 * 1024)}MB)",
+        flush=True,
+    )
+    # Model loading happened before this policy was known. Drop any free
+    # compilation/loading buffers it left behind so the first turn starts from
+    # the same bounded baseline as every later turn.
+    _reclaim_mlx_buffer_cache("startup-ready", force=True)
+
+
+_setup_gpu_memory_ceiling()
+
+
+def _augment_serial_stop_set(processor, model_dir: str) -> None:
+    """Merge generation_config.json's eos ids into the serial-path stop set.
+
+    The serial (non-batch) stream path stops on
+    `tokenizer.stopping_criteria.eos_token_ids`, seeded at load from the
+    tokenizer's own (sometimes narrower) eos set — the same gap that let
+    gemma4-e4b-q4 stream `<eos>` / `<|tool_response>` as text. We append
+    the declared ids directly (not via `add_eos_token_ids`, which expects
+    token *strings*). Defensive: mlx internals vary by version, so any
+    structural mismatch is logged and skipped rather than crashing
+    startup — the leaked-marker scrub still covers this path.
+    """
+    declared = _generation_config_eos_ids(model_dir) + _tool_call_close_stop_ids(model_dir)
+    if not declared:
+        return
+    try:
+        tok = getattr(processor, "tokenizer", None) or processor
+        criteria = getattr(tok, "stopping_criteria", None)
+        existing = getattr(criteria, "eos_token_ids", None)
+        if existing is None:
+            return
+        added = [e for e in declared if e not in existing]
+        existing.extend(added)
+        print(
+            f"[serial] stop set +{len(added)} from generation_config "
+            f"(now {len(existing)})",
+            flush=True,
+        )
+    except Exception as exc:  # noqa: BLE001 — never fail startup over this
+        print(f"[serial] stop-set augment skipped: {exc}", flush=True)
+
+
+_augment_serial_stop_set(PROCESSOR, ARGS.model)
+print("Application startup complete.", flush=True)
+print(
+    f"Uvicorn running on http://{ARGS.host}:{ARGS.port} (Press CTRL+C to quit)",
+    flush=True,
+)
+
+
+# ───────── Prompt cache layer ─────────
+
+# Per-cache_id state: (cache_state, est_bytes, last_used_ts).
+# The cache_state holds both the KV cache list and the token_ids
+# prefix; mlx-vlm uses these together to decide what to prefill.
+class CacheEntry:
+    __slots__ = ("state", "est_bytes", "last_used_ts", "dirty")
+
+    def __init__(
+        self,
+        state: PromptCacheState,
+        est_bytes: int,
+        last_used_ts: float,
+        dirty: bool = True,
+    ):
+        self.state = state
+        self.est_bytes = est_bytes
+        self.last_used_ts = last_used_ts
+        # True until this exact state is on disk. A shutdown flush writes only
+        # dirty entries: rewriting a 90k-token session that was loaded from
+        # disk unchanged is seconds of I/O, and the supervisor SIGKILLs three
+        # seconds after SIGTERM — measured, the one entry that mattered was
+        # still being written when the kill landed, and the restart re-prefilled
+        # 55k tokens.
+        self.dirty = dirty
+
+
+# Module-level cache. Bounded by `--cache-budget-mb` via LRU eviction.
+_CACHE: Dict[str, CacheEntry] = {}
+
+# Diagnostic (temporary, Phase-0 churn census): tool-roster shape recorded
+# at save time per cache_id so a later churn event can print BOTH sides'
+# shapes. Process-lifetime only — entries loaded from disk report
+# saved_roster=unknown, and the cached-text preview covers those.
+_LAST_SAVE_ROSTER: Dict[str, str] = {}
+
+
+def _roster_fp_of(tools) -> str:
+    """Compact fingerprint of a request's tool roster: "sha8/count", or "0"
+    for an empty roster. Diagnostics only — must never raise."""
+    try:
+        names = []
+        for t in tools or []:
+            if isinstance(t, dict):
+                fn = t.get("function") or {}
+                names.append(str(fn.get("name") or t.get("name") or "?"))
+            else:
+                fn = getattr(t, "function", None)
+                names.append(str(getattr(fn, "name", None) or getattr(t, "name", "?")))
+        if not names:
+            return "0"
+        digest = hashlib.sha256("\n".join(sorted(names)).encode("utf-8")).hexdigest()[:8]
+        return f"{digest}/{len(names)}"
+    except Exception:  # noqa: BLE001 — diagnostics never break a turn
+        return "err"
+
+
+# Cache ids (session + its seed prefix) currently being SERVED — registered
+# at the top of a chat-completions request and cleared when its stream ends.
+# Eviction (both in-memory `_enforce_budget` and `_enforce_disk_budget`) skips
+# these so the very entry a live turn is about to reuse can't be dropped out
+# from under it mid-flight. Without this pin, a long turn on a busy multi-
+# session box can have its own (least-recently-used, because it's been
+# streaming) prefix evicted, forcing a cold full re-prefill on the next step
+# — exactly the kind of churn that compounds the 27B first-token stall.
+# Multiset-counted (a value > 0 means pinned) so overlapping requests that
+# share a prefix id don't unpin each other early.
+_INFLIGHT_CACHE_IDS: "collections.Counter[str]" = collections.Counter()
+
+
+def _mark_inflight(*cache_ids: Optional[str]) -> List[str]:
+    """Pin the given (non-empty) cache ids from eviction. Returns the list
+    actually pinned, to hand to {@link _unmark_inflight} in a finally."""
+    pinned: List[str] = []
+    for cid in cache_ids:
+        if cid:
+            _INFLIGHT_CACHE_IDS[cid] += 1
+            pinned.append(cid)
+    return pinned
+
+
+def _unmark_inflight(cache_ids: List[str]) -> None:
+    # `.get` / `.pop(default)` rather than `[cid]` / `del`: a Counter returns 0
+    # for an absent key but `del` on it raises KeyError, so a stray double-
+    # release (or release of something never marked) must not crash the worker.
+    for cid in cache_ids:
+        count = _INFLIGHT_CACHE_IDS.get(cid, 0)
+        if count <= 1:
+            _INFLIGHT_CACHE_IDS.pop(cid, None)
+        else:
+            _INFLIGHT_CACHE_IDS[cid] = count - 1
+
+
+# Serializes the actual `stream_generate` GPU work across concurrent
+# requests. MLX text-only generation isn't safely batched in our wrapper
+# (no shared-prefix batching path, KV state lives on one Metal stream),
+# so two simultaneous chat completions would race the shared model
+# state and the per-cache_id KV save/restore. The TS-side ProviderQueue
+# now defaults to concurrency=2 to break the `ask_specialist`/`ask_gezel`
+# deadlock; this lock guarantees the engine still sees one stream at a
+# time. The lock is only held across the `for chunk in stream_generate(...)`
+# loop — cache lookup and cache save run outside it so the second request
+# can start its prefill prep while the first is still streaming.
+#
+# NOTE: must be constructed inside an event loop. FastAPI's lifespan
+# handles that — the lock is lazily allocated on first request.
+_GENERATION_LOCK: Optional[asyncio.Lock] = None
+
+
+def _get_generation_lock() -> asyncio.Lock:
+    global _GENERATION_LOCK
+    if _GENERATION_LOCK is None:
+        _GENERATION_LOCK = asyncio.Lock()
+    return _GENERATION_LOCK
+
+
+def _kv_quant_kwargs() -> Dict[str, Any]:
+    """KV-quantization options for the serial `stream_generate` path.
+
+    Forwarded through `stream_generate(**kwargs)` into mlx-vlm's
+    `generate_step`, which converts each layer via `to_quantized` once it
+    passes `quantized_kv_start`. Layer classes without that method
+    (BatchKVCache, ChunkedKVCache) are skipped by `maybe_quantize_kv_cache`
+    rather than raising — safe, but silent, which is why the launcher plans
+    f16 whenever the batched path is in use instead of trusting this flag.
+
+    Empty when disabled, so the default path is byte-identical to before.
+    """
+    if ARGS.kv_bits <= 0:
+        return {}
+    return {
+        "kv_bits": ARGS.kv_bits,
+        "kv_group_size": ARGS.kv_group_size,
+        "kv_quant_scheme": ARGS.kv_quant_scheme,
+        "quantized_kv_start": ARGS.quantized_kv_start,
+    }
+
+
+def _budget_bytes() -> int:
+    return ARGS.cache_budget_mb * 1024 * 1024
+
+
+# Fallback per-token byte estimate, used only when a state's layers don't
+# expose `nbytes` (every mlx_lm/mlx_vlm cache class does today, via
+# _BaseCache). The old blanket 100 KB/token figure measured ~55% high for
+# qwen3.8's 16 full-attention layers (64 KiB/token) and was blind to the
+# fixed ~147 MiB recurrent state of its 48 linear-attention layers; real
+# per-layer `nbytes` (shape math, no graph eval) prices both correctly.
+_BYTES_PER_TOKEN = 100 * 1024
+
+
+def _estimate_cache_bytes(state: PromptCacheState) -> int:
+    if state.token_ids is None:
+        return 0
+    layers = list(getattr(state, "cache", None) or [])
+    total = 0
+    complete = bool(layers)
+    try:
+        for layer in layers:
+            nb = getattr(layer, "nbytes", None)
+            if nb is None:
+                complete = False
+                break
+            total += int(nb)
+    except Exception:  # noqa: BLE001 — budget math must never break a turn
+        complete = False
+    if complete and total > 0:
+        return total
+    return len(state.token_ids) * _BYTES_PER_TOKEN
+
+
+def _enforce_budget() -> None:
+    """LRU eviction down to <= 80% of budget when over budget.
+
+    Two safety rails on top of the basic LRU:
+      1. NEVER evict the only entry. Evicting it buys nothing — the
+         next chat-completions request will immediately re-prefill it
+         back to the same size. Single-session users would loop
+         forever paying full prefill cost otherwise.
+      2. Stop evicting once one entry remains, regardless of budget.
+         An over-budget single entry is a signal that the budget is
+         too tight for this model; keeping the entry preserves cache
+         reuse across turns. Operators see usage exceeding budget
+         in /v1/cache/stats and can tune up.
+
+    When disk persistence is enabled, we save each evicted entry before
+    dropping it. The next session that asks for the same cache_id will
+    load it back from disk (cheap mmap read) instead of paying the
+    full prefill cost again.
+    """
+    total = sum(e.est_bytes for e in _CACHE.values())
+    budget = _budget_bytes()
+    if total < budget:
+        return
+    if len(_CACHE) <= 1:
+        return  # don't evict the only entry — cache reuse needs it
+    target = int(budget * 0.8)
+    # Sort by lru ascending; pop oldest until under target OR only one entry left.
+    items = sorted(_CACHE.items(), key=lambda kv: kv[1].last_used_ts)
+    for cache_id, entry in items:
+        if total < target or len(_CACHE) <= 1:
+            break
+        # Never evict an entry a live turn is currently serving — dropping
+        # it forces that turn into a cold re-prefill mid-flight.
+        if cache_id in _INFLIGHT_CACHE_IDS:
+            continue
+        # Persist before dropping. Best-effort — failure here just means
+        # the next session pays the prefill cost, which is the
+        # pre-persistence behavior. A clean entry is already on disk.
+        if _disk_enabled() and entry.dirty:
+            cache_persist.save_cache(
+                PERSIST_DIR, MODEL_FINGERPRINT, cache_id, entry.state  # type: ignore[arg-type]
+            )
+        del _CACHE[cache_id]
+        total -= entry.est_bytes
+        print(f"[cache] evicted cache_id={cache_id} (LRU, in-engine budget)", flush=True)
+    # Disk-budget enforcement runs independently of in-memory eviction:
+    # we may have just added new on-disk entries that push us over the
+    # disk cap. LRU prune by mtime keeps total disk usage bounded.
+    _enforce_disk_budget()
+
+
+def _enforce_disk_budget() -> None:
+    """Prune the oldest persisted cache files when over disk budget.
+
+    Off-by-default-when-zero: operators on a machine with abundant SSD
+    space can set --disk-cache-budget-mb=0 and let the OS reclaim
+    however much it needs. The default budget (8 GB) is generous for
+    the typical 25K-token Meester session (~2.5 GB) plus several
+    smaller sibling sessions.
+    """
+    if not _disk_enabled():
+        return
+    if ARGS.disk_cache_budget_mb <= 0:
+        return
+    budget_bytes = ARGS.disk_cache_budget_mb * 1024 * 1024
+    entries = cache_persist.list_persisted(PERSIST_DIR, MODEL_FINGERPRINT)  # type: ignore[arg-type]
+    total = sum(size for (_, size, _) in entries)
+    if total < budget_bytes:
+        return
+    target = int(budget_bytes * 0.8)
+    # Never prune the newest file. Same rail as the in-memory budget's "never
+    # evict the only entry": a single long session can outgrow the cap on its
+    # own (a 120k-token qwen3.8-27b snapshot is ~7.7 GB against the 8 GB
+    # default), and deleting the file a shutdown flush wrote a moment ago
+    # makes that flush — the only thing standing between a restart and a
+    # full re-prefill — pointless.
+    newest = entries[-1][0] if entries else None
+    for cache_id, size, _mtime in entries:
+        if total < target:
+            break
+        if cache_id == newest:
+            continue
+        # Skip the on-disk copy of an entry a live turn is serving — the
+        # in-memory pin above keeps it resident, but a sibling turn that
+        # disk-reloads the same prefix mid-flight must still find the file.
+        if cache_id in _INFLIGHT_CACHE_IDS:
+            continue
+        if cache_persist.delete_cache(PERSIST_DIR, MODEL_FINGERPRINT, cache_id):  # type: ignore[arg-type]
+            total -= size
+            print(f"[cache] disk-evicted cache_id={cache_id} (LRU, disk budget)", flush=True)
+
+
+def _save_cache(cache_id: str, state: PromptCacheState) -> None:
+    bytes_est = _estimate_cache_bytes(state)
+    _CACHE[cache_id] = CacheEntry(state=state, est_bytes=bytes_est, last_used_ts=time.time())
+    _enforce_budget()
+
+
+def _try_load_from_disk(cache_id: str) -> Optional[PromptCacheState]:
+    """Pull a cache entry off disk into _CACHE under the given id.
+
+    Returns the loaded state on hit (also installs in _CACHE), None on
+    miss / corruption / fingerprint mismatch. The next save will write
+    back through the normal path.
+    """
+    if not _disk_enabled():
+        return None
+    state = cache_persist.load_cache(PERSIST_DIR, MODEL_FINGERPRINT, cache_id)  # type: ignore[arg-type]
+    if state is None:
+        return None
+    bytes_est = _estimate_cache_bytes(state)
+    _CACHE[cache_id] = CacheEntry(
+        state=state, est_bytes=bytes_est, last_used_ts=time.time(), dirty=False
+    )
+    print(
+        f"[cache] disk-load cache_id={cache_id} tokens={len(state.token_ids or [])}",
+        flush=True,
+    )
+    return state
+
+
+def _try_seed_from_prefix(cache_id: str, prefix_cache_id: str) -> Optional[PromptCacheState]:
+    """When a session misses, fall back to its gezel-prefix entry.
+
+    Loads the prefix-shaped entry from disk and registers it under
+    `cache_id` so the session inherits it as its starting state. The
+    prefix file is never overwritten by the session — its own delta
+    saves under its own cache_id. This is how new sessions for an
+    existing gezel skip the system-prompt prefill cost.
+
+    Returns None on any miss; the caller falls through to fresh state.
+    """
+    if not _disk_enabled():
+        return None
+    state = cache_persist.load_cache(
+        PERSIST_DIR, MODEL_FINGERPRINT, prefix_cache_id  # type: ignore[arg-type]
+    )
+    if state is None:
+        return None
+    bytes_est = _estimate_cache_bytes(state)
+    # Clean: this is the prefix file's content, already on disk under the
+    # prefix id. Persisting a copy under the session id buys nothing.
+    _CACHE[cache_id] = CacheEntry(
+        state=state, est_bytes=bytes_est, last_used_ts=time.time(), dirty=False
+    )
+    print(
+        f"[cache] prefix-seed cache_id={cache_id} from prefix={prefix_cache_id} "
+        f"tokens={len(state.token_ids or [])}",
+        flush=True,
+    )
+    return state
+
+
+def _request_prefix_ids(request) -> List[str]:
+    """Ordered, most-specific-first prefix ids for this request. Prefers the
+    layered `prefix_cache_ids`; falls back to the singular `prefix_cache_id`
+    for back-compat. Filters out empties."""
+    if getattr(request, "prefix_cache_ids", None):
+        return [p for p in request.prefix_cache_ids if p]
+    if request.prefix_cache_id:
+        return [request.prefix_cache_id]
+    return []
+
+
+# Prefix ids we've already seeded from a real session save this process.
+# Mirrors the llama-cpp adapter's `seededPrefixes` — seed once per process
+# so back-to-back sessions don't repeatedly rewrite the same prefix file.
+_SEEDED_PREFIXES: set = set()
+
+
+def _band_snapshot_target(sub) -> Optional[int]:
+    """Token index of the shared band's end for this request, or None.
+
+    The TS side knows the band as a CHARACTER length inside the system
+    message (`stable_prefix_chars`); only the engine has the tokenizer, and
+    the rendered prompt carries template framing plus the tool block ahead
+    of the system content — so the offset is located by finding the first
+    token index whose decode contains the band's tail. See ADR 0010.
+    """
+    n_chars = getattr(sub.request, "stable_prefix_chars", None)
+    if not n_chars or int(n_chars) <= 0:
+        return None
+    system = next(
+        (m for m in (getattr(sub.request, "messages", None) or [])
+         if getattr(m, "role", None) == "system"),
+        None,
+    )
+    content = getattr(system, "content", None) if system is not None else None
+    if not isinstance(content, str):
+        return None
+    marker = content[: int(n_chars)][-_BAND_MARKER_CHARS:]
+    if len(marker) < _BAND_MARKER_MIN_CHARS:
+        return None
+    tok = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+    toks = list(sub.prompt_tokens or [])
+    try:
+        target = cache_seed.token_boundary_for_marker(
+            lambda k: tok.decode(toks[:k]), len(toks), marker, _SNAPSHOT_BOUNDARY_MARGIN
+        )
+    except Exception as exc:  # noqa: BLE001 — best-effort; fall back to today
+        print(f"[cache] band-boundary probe failed: {exc}", flush=True)
+        return None
+    # One line naming the decision — the thing to grep when a sibling
+    # re-prefills instead of extending. `target=0` means the band's tail was
+    # not found in the render, which is a marker problem, not a cache one.
+    print(
+        f"[cache] band-boundary chars={int(n_chars)} target={target} "
+        f"prompt_tokens={len(toks)}",
+        flush=True,
+    )
+    return target or None
+
+
+def _seed_prefix_from_session(request, state: PromptCacheState) -> None:
+    """Seed the most-specific (gp) prefix entry from a real session's saved
+    state so a SIBLING session (same gezel+project, different volatile state)
+    inherits the warm `[system+tools]` prefix on its next turn — the
+    cross-session win the layered flag exists for.
+
+    Mirrors the llama-cpp adapter, which copies a slot save into a
+    `prefix-<hash>.bin` file. The seeded entry holds this session's full
+    state (system+tools+transcript); a sibling's longest-common-prefix trim
+    reuses only the shared `[system+tools]` head and re-prefills its own
+    divergent tail. Best-effort, once per prefix per process; never blocks
+    the request path on failure.
+    """
+    if not _disk_enabled():
+        return
+    ids = _request_prefix_ids(request)
+    if not ids:
+        return
+    gp = ids[0]
+    if gp in _SEEDED_PREFIXES:
+        return
+    try:
+        # INVARIANT: a prefix entry may never be replaced by a LONGER state.
+        # On this model family `seed_from_state` reuses an entry only when it
+        # is a strict prefix (`lcp == n` -> extension); an entry longer than
+        # the shared head cannot be trimmed (48 of 64 layers are
+        # linear-attention `ArraysCache`), so it degrades every sibling to a
+        # full re-prefill. Lengthening is therefore never an improvement, and
+        # a shorter entry that already works must survive a later process
+        # that would otherwise overwrite it with a whole session. Measured:
+        # a 7,202-token entry that delivered 89.8% reuse was overwritten by
+        # the same session's 8,008-token post-turn state moments later.
+        incoming = len(state.token_ids or [])
+        existing = cache_persist.load_cache(PERSIST_DIR, MODEL_FINGERPRINT, gp)
+        prior = len(getattr(existing, "token_ids", None) or []) if existing else 0
+        if prior and incoming > prior:
+            _SEEDED_PREFIXES.add(gp)
+            print(
+                f"[cache] prefix-keep {gp}: existing {prior} tokens is shorter "
+                f"than this session's {incoming} — a longer entry cannot be "
+                f"trimmed by a sibling",
+                flush=True,
+            )
+            return
+        if cache_persist.save_cache(PERSIST_DIR, MODEL_FINGERPRINT, gp, state):  # type: ignore[arg-type]
+            _SEEDED_PREFIXES.add(gp)
+            _LAST_SAVE_ROSTER[gp] = _roster_fp_of(getattr(request, "tools", None))
+            _enforce_disk_budget()
+            print(
+                f"[cache] prefix-seeded {gp} from cache_id={request.cache_id} "
+                f"tokens={len(state.token_ids or [])}",
+                flush=True,
+            )
+    except Exception as exc:  # noqa: BLE001 — best-effort, never fail the turn
+        print(f"[cache] prefix-seed failed for {gp}: {exc}", flush=True)
+
+
+# ───────── BatchGenerator engine (--max-concurrency >= 1) ─────────
+#
+# Any positive concurrency routes through BatchEngine instead of the legacy
+# serial `stream_generate` + _GENERATION_LOCK path. At N=1 this is still a
+# singleton queue, but it matters for correctness/performance: BatchGenerator
+# exposes prompt-segment boundaries, which lets windowed Qwen/Gemma caches save
+# a coherent prompt snapshot instead of cold-prefilling after every tool-loop
+# turn. At N>1 the same engine also continuously batches independent sequences.
+# One asyncio worker drives the generator on the event-loop thread — MLX GPU
+# streams are thread-local (see the chat_completions note). Every per-session
+# concern maps to a per-sequence list in the BatchGenerator API (sampler, logits
+# processors incl. the tool grammar, seed KV cache, max_tokens); on finish we
+# extract the sequence's KV and persist it via the SAME _save_cache /
+# _seed_prefix_from_session path the serial route uses.
+
+
+def _resolve_cache_state(request):
+    """The cache cascade (memory → disk → prefix-seed → fresh), shared by
+    the serial and batched paths. Returns (state, source, matched_prefix)."""
+    cache_state = None
+    cache_source = "fresh"
+    matched_prefix = None
+    if request.cache_id:
+        existing_entry = _CACHE.get(request.cache_id)
+        if existing_entry is not None:
+            cache_state = existing_entry.state
+            existing_entry.last_used_ts = time.time()
+            cache_source = "memory"
+        else:
+            disk_state = _try_load_from_disk(request.cache_id)
+            if disk_state is not None:
+                cache_state = disk_state
+                cache_source = "disk"
+            else:
+                for pid in _request_prefix_ids(request):
+                    seeded = _try_seed_from_prefix(request.cache_id, pid)
+                    if seeded is not None:
+                        cache_state = seeded
+                        cache_source = "prefix"
+                        matched_prefix = pid
+                        break
+    if cache_state is None:
+        cache_state = PromptCacheState()
+    return cache_state, cache_source, matched_prefix
+
+
+class _UnwrapLM:
+    """mlx_vlm's LanguageModel returns LanguageModelOutput(.logits); mlx_lm's
+    BatchGenerator wants a raw logits array. Thin shim: unwrap `.logits` and
+    delegate every other attribute (layers, config, make_cache, …) to the
+    wrapped model."""
+
+    def __init__(self, inner):
+        self._inner = inner
+        self.layers = inner.layers
+
+    def __call__(self, *args, **kwargs):
+        out = self._inner(*args, **kwargs)
+        return getattr(out, "logits", out)
+
+    def __getattr__(self, name):
+        return getattr(self._inner, name)
+
+
+# Batched-prefill liveness tuning. Only emit a marker for waves big enough
+# that prefill is non-trivial (small prompts return a first token fast, so a
+# marker would just be noise), and no more often than the interval so a
+# chunk-iterating worker loop doesn't spam the log.
+_PREFILL_LIVENESS_MIN_TOKENS = 2048
+
+# Event-loop breathing between engine steps — see BatchEngine._breathe.
+_BREATHE_EVERY_S = 0.2
+_BREATHE_FOR_S = 0.002
+_PREFILL_LIVENESS_INTERVAL_S = 4.0
+
+# Keep chunk-edge snapshots this many tokens clear of the prompt's end.
+# The final tokens (`…<start_of_turn>model\n`) re-merge differently once
+# the next turn's text continues past them (BPE boundary instability), so
+# a snapshot cut too close to the end misses by a couple of tokens — and
+# wrapped rotating layers can't trim even that. 16 is generous; the cost
+# is at most 16 re-prefilled tokens per turn.
+_SNAPSHOT_BOUNDARY_MARGIN = 16
+# Tail of the shared band used to locate its token boundary. Long enough to
+# be unique in a 60k-char prompt, short enough that one BPE re-merge at the
+# cut cannot swallow it.
+_BAND_MARKER_CHARS = 256
+_BAND_MARKER_MIN_CHARS = 64
+
+
+def _prefix_seed_state(sub, session_state, trimmable: bool):
+    """The state to publish as this request's prefix entry, or None.
+
+    A band-keyed entry (ADR 0010) must be the band itself. A longer
+    untrimmable entry is not a partial hit for a sibling but a full
+    re-prefill, so when the band snapshot is missing — never planted,
+    or dropped by the boundary proof — nothing is published. A trimmable
+    stack keeps publishing its post-turn state: siblings trim it back.
+    """
+    if sub.band_snapshot is not None:
+        layers, count = sub.band_snapshot
+        state = PromptCacheState()
+        state.cache = layers
+        state.token_ids = list(sub.prompt_tokens[:count])
+        return state
+    if getattr(sub.request, "stable_prefix_chars", None) and not trimmable:
+        return None
+    return session_state
+
+
+def _store_snapshot(sub, snapshot) -> None:
+    """File a verified boundary snapshot under the slot(s) its boundary is:
+    the session's own save, the shared band, or both when they coincide."""
+    at_tokens = snapshot[1]
+    if at_tokens == sub.snapshot_target:
+        sub.prompt_snapshot = snapshot
+    if at_tokens == sub.band_target:
+        sub.band_snapshot = snapshot
+
+
+class _Sub:
+    """One in-flight batched request: its decode queue, accumulated tokens,
+    and the bits needed to build its sequence + persist its KV on finish."""
+
+    __slots__ = (
+        "request", "prompt_tokens", "grammar", "seed_state", "request_id",
+        "http_request", "uid", "queue", "cancelled", "token_ids", "emitted",
+        "pinned_cache_ids", "prefill_total", "first_token_seen",
+        "reused_tokens", "seed_mode", "prompt_snapshot", "snapshot_target",
+        "band_target", "band_snapshot", "saved_cache_state",
+        "prefill_started_at", "generation_started_at",
+        "prompt_tps", "generation_tps", "roster_fp",
+        "priority", "queued_at",
+    )
+
+    def __init__(
+        self,
+        request,
+        prompt_tokens,
+        grammar,
+        seed_state,
+        request_id,
+        http_request,
+        snapshot_target=None,
+        roster_fp=None,
+    ):
+        self.request = request
+        self.prompt_tokens = prompt_tokens   # List[int], full prompt
+        self.grammar = grammar               # SafeToolGrammarProcessor | None
+        self.seed_state = seed_state         # PromptCacheState from the cascade
+        self.request_id = request_id
+        self.http_request = http_request
+        # Diagnostic (temporary, Phase-0 churn census): the tool-roster
+        # shape this sub's prompt was rendered with. The warm path passes
+        # it explicitly because its internal ChatRequest omits tools.
+        self.roster_fp = (
+            roster_fp
+            if roster_fp is not None
+            else _roster_fp_of(getattr(request, "tools", None))
+        )
+        self.uid = None
+        self.queue: asyncio.Queue = asyncio.Queue()
+        self.cancelled = False
+        self.token_ids = []                  # generated token ids
+        self.emitted = 0                     # chars already streamed
+        self.pinned_cache_ids: List[str] = []  # eviction pins to release on stream end
+        self.prefill_total = 0               # new tokens this sub must prefill
+        self.first_token_seen = False        # flips once decode emits a token
+        self.reused_tokens = 0               # cache tokens actually served (seed plan)
+        self.seed_mode = "fresh"             # seed decision, for logs/diagnosis
+        # End-of-prompt KV snapshot (windowed models only) — captured at
+        # first token, before any generated token enters the cache, so
+        # the saved session entry stays a pure prefix of the next turn's
+        # re-templated prompt. See cache_seed.probe_needs_prompt_snapshot.
+        self.prompt_snapshot = None
+        # Absolute token boundary to capture. Normal turns default to
+        # len(prompt)-margin at admission; prefix warming supplies the stable
+        # boundary found by comparing two different synthetic user turns.
+        self.snapshot_target = snapshot_target
+        # Shared-band boundary (ADR 0010) and its snapshot, captured beside
+        # the session's own so a fresh turn can publish the short band as the
+        # sibling prefix entry without shrinking its own saved cache to it.
+        self.band_target = None
+        self.band_snapshot = None
+        # The exact state committed at finish. Cache warming uses this direct
+        # reference for immediate persistence even if the memory-budget pass
+        # evicts the just-built entry from the global LRU.
+        self.saved_cache_state = None
+        # BatchGenerator does not expose per-sequence timing. Track wall-clock
+        # timing per subscription so the OpenAI-compatible usage frames retain
+        # the same prompt/decode speed fields as the serial stream_generate
+        # path. Values are snapshotted into queue items because the worker can
+        # advance again before the SSE coroutine consumes them.
+        self.prefill_started_at = None
+        self.generation_started_at = None
+        self.prompt_tps = 0.0
+        self.generation_tps = 0.0
+        # Who is waiting on this request: "interactive" (a person) or
+        # "background" (task steps, warms, chores). Decides wave membership
+        # and whether a running wave steps aside — see wave_policy.py.
+        self.priority = wave_policy.normalize_priority(getattr(request, "priority", None))
+        self.queued_at = time.time()
+
+
+class BatchEngine:
+    """Owns one mlx_lm BatchGenerator + the single worker that drives it."""
+
+    def __init__(self, model, tokenizer, max_concurrency):
+        from mlx_lm.generate import BatchGenerator  # local: only when batching
+
+        self._tok = tokenizer
+        self._max = max(1, int(max_concurrency))
+        eos = getattr(tokenizer, "eos_token_ids", None)
+        if eos is None:
+            eos = getattr(tokenizer, "eos_token_id", None)
+        if eos is None:
+            eos_list = []
+        elif isinstance(eos, int):
+            eos_list = [eos]
+        else:
+            eos_list = list(eos)
+        # Union in the stop set the model author declared in
+        # generation_config.json — the tokenizer wrapper sometimes
+        # resolves a narrower eos set than the model actually ends turns
+        # on (see _generation_config_eos_ids).
+        eos_from_cfg = _generation_config_eos_ids(ARGS.model)
+        for e in eos_from_cfg:
+            if e not in eos_list:
+                eos_list.append(e)
+        # Halt at the first closed `<tool_call>` for models that declare a
+        # server-side tool_call_parser but don't list `</tool_call>` in eos
+        # (poolside Laguna) — otherwise the raw completion path never stops
+        # after a call and loops hundreds of them per turn until the SSE
+        # stalls. See _tool_call_close_stop_ids.
+        tool_close = _tool_call_close_stop_ids(ARGS.model)
+        for e in tool_close:
+            if e not in eos_list:
+                eos_list.append(e)
+        self._stop_ids = {int(e) for e in eos_list}
+        stop_tokens = [[int(e)] for e in eos_list]
+        self._model = model
+        self._gen_kwargs = dict(
+            max_tokens=2048,
+            stop_tokens=stop_tokens or None,
+            completion_batch_size=self._max,
+            prefill_batch_size=self._max,
+            prefill_step_size=ARGS.prefill_step_size,
+        )
+        self._gen = self._make_generator("batch-init")
+        self._subs = {}            # uid -> _Sub
+        self._pending = []         # _Sub awaiting insert
+        self._wake = asyncio.Event()
+        self._worker = None
+        # Preemption state (wave_policy.py). A wave made only of background
+        # work parks while waiting interactive requests run on a second
+        # generator; the parked wave keeps its KV and resumes where it stopped.
+        self._spare_gen = None     # second BatchGenerator, built on first preemption
+        self._nested = False       # True while an interactive wave runs over a parked one
+        self._parked = []          # subs of the parked wave (paused markers)
+        self._held_background = [] # background arrivals while nested
+        self._spec_sub = None      # the sub a spec wave is serving
+        # Set while a spec wave's rounds own the drafter's per-wave state
+        # (its cache, shared kv, position); a wave run inside the pause must
+        # not draft with the same drafter.
+        self._drafter_busy = False
+        self._liveness = wave_policy.LivenessThrottle()
+        self._preempt_memory_noted = False
+        self._last_breath = 0.0
+        # Windowed / state-cache models can't rewind their caches once
+        # the sequence outgrows the window, so a post-generation save is
+        # dead weight the moment history diverges from the raw tokens
+        # (reasoning stripping guarantees it does). For those models we
+        # capture an end-of-prompt KV snapshot per turn instead — see
+        # _maybe_snapshot_prompt_cache. Full-attention stacks skip the
+        # copy; LCP+trim covers them.
+        self._needs_snapshot = cache_seed.probe_needs_prompt_snapshot(model)
+        self._snapshot_warned = False
+        # Prefill-liveness for the current wave. The batched path prefills the
+        # whole prompt inside BatchGenerator with NO tqdm bar (unlike the
+        # serial mlx_vlm path), so the TS watchdog saw zero prefill events and
+        # the engine looked silent through a long first-token wait (the
+        # qwen3.6-27b 37K-token stall). We emit a `Prefill:`-shaped marker the
+        # stdout-parser already understands so the watchdog knows the engine is
+        # alive + prefilling N tokens, and the UI pill shows it.
+        self._prefill_total = 0      # new tokens the active wave is prefilling
+        self._prefill_done = {}      # uid → tokens prefilled so far (this wave)
+        # uid → (cache_id, that sub's own prefill total). The engine has ONE
+        # stdout, so an untagged `Prefill:` line is broadcast by the supervisor
+        # to every in-flight session — a session streaming its reply would draw
+        # a neighbour's prefill bar on its own row, at that neighbour's token
+        # count. This map is what makes each marker addressable.
+        self._prefill_meta = {}
+        self._last_prefill_emit = 0.0
+        self._memory_check_steps = 0
+        print(
+            f"[batch] engine ready max_concurrency={self._max} "
+            f"stop_tokens={len(stop_tokens)} (+{len(eos_from_cfg)} from generation_config"
+            f"{f', +{len(tool_close)} </tool_call>' if tool_close else ''}) "
+            f"prompt_snapshot={'y' if self._needs_snapshot else 'n'}",
+            flush=True,
+        )
+
+    def _make_generator(self, phase):
+        from mlx_lm.generate import BatchGenerator  # local: only when batching
+
+        gen = BatchGenerator(self._model, **self._gen_kwargs)
+        # mlx_lm BatchGenerator sets the wired limit to Metal's entire
+        # recommended working set in its constructor. Restore the smaller
+        # resident-set plan and free-buffer cap selected by Gezel.
+        _apply_mlx_memory_policy(phase)
+        return gen
+
+    async def _breathe(self):
+        """Yield to the event loop between engine steps.
+
+        `sleep(0)` yields exactly one loop iteration, and a NEW request needs
+        several (accept, parse, route, read body, render, tokenize, submit)
+        before the engine ever sees it. Each engine step blocks the loop for
+        its whole duration — a 2048-token prefill chunk is seconds on a 27B —
+        so one-iteration yields made an arriving request pay that many steps
+        before it could even queue. Measured: 8.0 s to first token for a
+        37-token question whose engine-side wait was 0.0 s. A short real
+        sleep every ~200 ms lets the loop drain all ready I/O in one go, at
+        a cost bounded by the ratio (~1%) rather than per token."""
+        now = time.perf_counter()
+        if now - self._last_breath >= _BREATHE_EVERY_S:
+            self._last_breath = now
+            await asyncio.sleep(_BREATHE_FOR_S)
+        else:
+            await asyncio.sleep(0)
+
+    def submit(self, sub):
+        if self._worker is None:
+            self._worker = asyncio.ensure_future(self._run())
+        if self._nested and wave_policy.is_background(sub):
+            # An interactive wave is running over a parked one; background
+            # arrivals wait until the parked wave has resumed.
+            self._held_background.append(sub)
+        else:
+            self._pending.append(sub)
+        self._announce_queued(sub)
+        self._wake.set()
+
+    def is_idle(self):
+        """True once no live or queued batch sub can still use scratch buffers."""
+        return (
+            not self._subs
+            and not self._pending
+            and not self._held_background
+            and not self._parked
+            and self._spec_sub is None
+        )
+
+    def _running_subs(self):
+        running = list(self._subs.values())
+        if self._spec_sub is not None:
+            running.append(self._spec_sub)
+        return running
+
+    def _wave_running(self):
+        return bool(self._subs) or self._spec_sub is not None or self._nested
+
+    def _announce_queued(self, sub):
+        """One line naming a request that arrived mid-wave, plus its first
+        waiting marker right away — the TS side flips the turn from
+        "Processing prompt" to "Waiting for …" without waiting for the next
+        worker step, which on a long prefill chunk can be most of a minute."""
+        if not self._wave_running():
+            return
+        cid = getattr(sub.request, "cache_id", None)
+        behind = wave_policy.cache_ids(self._running_subs())
+        ahead = max(0, len(self._pending) + len(self._held_background) - 1)
+        print(
+            f"[batch] queued request={sub.request_id} cache={cid or '-'} "
+            f"priority={sub.priority} ahead={ahead} behind={','.join(behind) or '-'}",
+            flush=True,
+        )
+        if cid and self._liveness.due(cid):
+            print(wave_policy.waiting_line(cid, ahead, behind), flush=True)
+
+    def _emit_waiting_liveness(self):
+        """Liveness for requests the worker is NOT serving right now: queued
+        ones (`waiting`) and a parked background wave (`paused`). Printed from
+        the worker loop, between steps — so, like the prefill marker, it is
+        proof the engine is alive and busy, and a wedged worker goes quiet.
+
+        Without it a queued request's only signal was silence: the TS
+        pre-first-byte watchdog aborted a turn that was waiting its turn, and
+        a parked wave's open stream would trip the streaming-idle bound."""
+        running = self._running_subs()
+        if not running:
+            return
+        behind = wave_policy.cache_ids(running)
+        queued = [s for s in self._pending + self._held_background if wave_policy.is_live(s)]
+        for idx, sub in enumerate(queued):
+            cid = getattr(sub.request, "cache_id", None)
+            if cid and self._liveness.due(cid):
+                print(wave_policy.waiting_line(cid, idx, behind), flush=True)
+        for sub in self._parked:
+            cid = getattr(sub.request, "cache_id", None)
+            if cid and wave_policy.is_live(sub) and self._liveness.due(cid):
+                print(wave_policy.paused_line(cid, behind), flush=True)
+
+    def _announce_admitted(self, sub):
+        cid = getattr(sub.request, "cache_id", None)
+        if cid and self._liveness.seen(cid):
+            print(wave_policy.admitted_line(cid, time.time() - sub.queued_at), flush=True)
+        self._liveness.forget(cid)
+
+    def _drop_cancelled_pending(self):
+        """A request whose client left while it was queued must not be
+        admitted. Measured: a turn the watchdog had already abandoned was
+        admitted two minutes later and prefilled for nobody."""
+        live = []
+        for sub in self._pending:
+            if sub.cancelled:
+                print(
+                    f"[batch] dropped request={sub.request_id} "
+                    f"cache={getattr(sub.request, 'cache_id', None) or '-'} "
+                    f"(client left while queued)",
+                    flush=True,
+                )
+                self._liveness.forget(getattr(sub.request, "cache_id", None))
+                continue
+            live.append(sub)
+        self._pending = live
+
+    def _memory_allows_preemption(self):
+        """A parked wave keeps its KV resident while the interactive wave
+        allocates its own, so preempt only with headroom below the same
+        threshold that collapses a wave to one member."""
+        if _MEM_LIMIT_BYTES <= 0:
+            return True
+        active = _safe_active_memory()
+        if active <= 0:
+            return True
+        ok = active < _MEM_LIMIT_BYTES * _MEM_ADMIT_FRACTION
+        if not ok and not self._preempt_memory_noted:
+            self._preempt_memory_noted = True
+            print(
+                f"[batch] preempt held: active={active // (1024 * 1024)}MB is at the "
+                f"admission threshold of ceiling={_MEM_LIMIT_BYTES // (1024 * 1024)}MB; "
+                f"the waiting request runs after this wave",
+                flush=True,
+            )
+        return ok
+
+    def _preempt_due(self, running):
+        if self._nested:
+            return False
+        if not wave_policy.preemptors(running, self._pending):
+            return False
+        return self._memory_allows_preemption()
+
+    async def _serve_preemptors(self, parked):
+        """Run waiting interactive requests to completion while `parked` (a
+        wave of background work) holds still, then hand the engine back.
+
+        The parked wave keeps everything it owns — KV, generator, logits
+        processor state, its open SSE stream — and resumes at the step it
+        stopped on. The interactive wave runs on a second BatchGenerator so
+        the parked one's batch state is untouched; the shared per-wave fields
+        (prefill liveness, the spec sub) are saved and restored around it.
+        Nothing is cancelled, so nothing is recomputed.
+
+        Safe points only: callers park between engine steps (a batched
+        wave's `next()`, a spec wave's prefill chunk or committed round), so
+        no verification transaction or half-written cache is ever left open.
+        """
+        fg = [s for s in self._pending if wave_policy.is_live(s) and not wave_policy.is_background(s)]
+        if not fg:
+            return
+        if self._spare_gen is None:
+            try:
+                self._spare_gen = self._make_generator("batch-preempt-init")
+            except Exception as exc:  # noqa: BLE001 — never kill the worker over a preemption
+                print(f"[batch] preempt unavailable (spare generator failed): {exc}", flush=True)
+                log_contained_exception("batch")
+                return
+        held = [s for s in self._pending if s not in fg]
+        parked_ids = wave_policy.cache_ids(parked)
+        print(
+            f"[batch] preempt: parking {len(parked)} background request(s) "
+            f"[{','.join(parked_ids) or '-'}] for {len(fg)} interactive "
+            f"[{','.join(wave_policy.cache_ids(fg)) or '-'}]",
+            flush=True,
+        )
+        started = time.perf_counter()
+        saved = (
+            self._gen,
+            self._subs,
+            self._prefill_total,
+            self._prefill_done,
+            self._prefill_meta,
+            self._spec_sub,
+        )
+        self._gen = self._spare_gen
+        self._subs = {}
+        self._prefill_total = 0
+        self._prefill_done = {}
+        self._prefill_meta = {}
+        self._last_prefill_emit = 0.0
+        self._spec_sub = None
+        self._pending = fg
+        self._held_background = held
+        self._parked = list(parked)
+        self._nested = True
+        try:
+            while self._pending or self._subs:
+                await self._step_wave()
+        finally:
+            self._spare_gen = self._gen
+            (
+                self._gen,
+                self._subs,
+                self._prefill_total,
+                self._prefill_done,
+                self._prefill_meta,
+                self._spec_sub,
+            ) = saved
+            self._last_prefill_emit = 0.0
+            self._pending = self._pending + self._held_background
+            self._held_background = []
+            self._parked = []
+            self._nested = False
+            self._preempt_memory_noted = False
+            for sub in parked:
+                self._liveness.forget(getattr(sub.request, "cache_id", None))
+        print(
+            f"[batch] resume: {len(parked)} background request(s) "
+            f"[{','.join(parked_ids) or '-'}] after {time.perf_counter() - started:.1f}s",
+            flush=True,
+        )
+
+    def _save_cancelled_snapshot(self, sub, cache_layers=None, wave_tokens=None, done_tokens=None):
+        """Keep what a cancelled request already paid for.
+
+        A cancel — the user's stop, a watchdog, or the service shutting down —
+        used to drop the turn's prompt KV on the floor: the end-of-prompt
+        snapshot is filed only by `_finish`, which a cancelled wave never
+        reaches. The restart that motivated this lost a 98k-token prompt that
+        had been fully prefilled two minutes earlier, and the resumed turn
+        re-prefilled from a 14k-token prefix.
+
+        Saves the planted end-of-prompt snapshot when it was captured, or —
+        for a spec wave stopped mid-prefill — the prefix it reached, proven
+        against the prompt like any other snapshot. Only ever an improvement:
+        it must extend past what this turn was seeded with, so a retry of the
+        same turn resumes from here instead of from the seed."""
+        cid = getattr(sub.request, "cache_id", None)
+        if not cid or not self._needs_snapshot:
+            return
+        try:
+            snapshot = sub.prompt_snapshot
+            if snapshot is None and cache_layers is not None and wave_tokens and done_tokens:
+                boundary = int(sub.reused_tokens) + int(done_tokens)
+                layers = spec_decode.clone_layers(cache_layers)
+                if cache_seed.snapshot_matches_prompt(
+                    layers, list(wave_tokens)[:boundary], sub.prompt_tokens, boundary
+                ):
+                    snapshot = (layers, boundary)
+            if snapshot is None:
+                return
+            layers, count = snapshot
+            if int(count) <= int(sub.reused_tokens or 0):
+                return
+            state = PromptCacheState()
+            state.cache = layers
+            state.token_ids = list(sub.prompt_tokens[: int(count)])
+            _save_cache(cid, state)
+            _LAST_SAVE_ROSTER[cid] = sub.roster_fp
+            print(
+                f"[batch] saved cancelled cache_id={cid} tokens={int(count)} "
+                f"(was seeded with {int(sub.reused_tokens or 0)})",
+                flush=True,
+            )
+        except Exception as exc:  # noqa: BLE001 — best-effort; never break the worker
+            print(f"[batch] cancelled-turn save failed cache_id={cid}: {exc}", flush=True)
+
+    def _sampler_for(self, request):
+        from mlx_lm.sample_utils import make_sampler
+
+        return make_sampler(
+            temp=float(request.temperature) if request.temperature is not None else 0.0,
+            top_p=float(request.top_p) if request.top_p is not None else 0.0,
+            top_k=int(request.top_k) if request.top_k is not None else 0,
+        )
+
+    def _processors_for(self, request, grammar):
+        from mlx_lm.sample_utils import make_logits_processors
+
+        procs = []
+        if request.repetition_penalty is not None:
+            procs = make_logits_processors(
+                repetition_penalty=float(request.repetition_penalty),
+                repetition_context_size=int(request.repetition_context_size or 20),
+            )
+        if grammar is not None:
+            # Accepts a single processor or a list (grammar + think-budget).
+            procs = list(procs) + (grammar if isinstance(grammar, list) else [grammar])
+        # BatchGenerator receives one processor list per sequence. A mixed
+        # wave (for example cache warm + pi tool call) must use [] for the
+        # unconstrained sequence: mlx_lm iterates every per-sequence entry and
+        # crashes on None with "'NoneType' object is not iterable".
+        return list(procs or [])
+
+    def _seed_args(self, sub):
+        """Plan cache reuse for this sub: longest-common-prefix with
+        trim-to-prefix, mirroring the serial path's semantics (the old
+        exact-prefix-only check re-prefilled entire 20K-token sessions
+        whenever one stripped reasoning character diverged the history
+        from the raw tokens). Delegates to cache_seed.seed_from_state;
+        stamps the authoritative prefill/reuse counts on the sub so the
+        liveness marker and usage frames report engine reality, and logs
+        the decision — the line to grep when turns re-prefill."""
+        # Late prefix re-check. The lookup in the request handler runs when the
+        # HTTP request ARRIVES; a band is published when its pioneer's turn
+        # ENDS. Static-wave admission puts minutes between the two, so every
+        # sibling dispatched alongside the pioneer — which is the whole fanout
+        # case this exists for — resolved `fresh` before the entry existed.
+        # Measured: two real sibling sessions agreed on
+        # `prefix-band-8977c4fc5f74e5d9`, the pioneer published 16,576 tokens
+        # to it, and the sibling still cold-prefilled 71,346 because its
+        # lookup had run 55ms too early. Re-checking here, at the moment the
+        # wave actually starts, is the only point that sees the publish.
+        if sub.request.cache_id and not getattr(sub.seed_state, "token_ids", None):
+            for pid in _request_prefix_ids(sub.request):
+                seeded = _try_seed_from_prefix(sub.request.cache_id, pid)
+                if seeded is not None:
+                    sub.seed_state = seeded
+                    break
+        plan = cache_seed.seed_from_state(sub.seed_state, sub.prompt_tokens)
+        # A session entry that diverged and cannot be trimmed is worth
+        # nothing to this prompt, but the shared prefix it was born from
+        # usually still is. Without this a stale session entry was strictly
+        # WORSE than having none: a miss falls back to the prefix band
+        # (~14-17k tokens reused on qwen3.8-27b), a divergent hit re-prefilled
+        # everything. It matters most right after a restart, when the session
+        # entry the engine persisted no longer matches a rebuilt transcript.
+        if sub.request.cache_id and str(plan.mode).startswith("fresh") and plan.cached_len:
+            for pid in _request_prefix_ids(sub.request):
+                seeded = _try_seed_from_prefix(sub.request.cache_id, pid)
+                if seeded is None:
+                    continue
+                fallback = cache_seed.seed_from_state(seeded, sub.prompt_tokens)
+                if fallback.reused > plan.reused:
+                    print(
+                        f"[batch] seed fallback cache_id={sub.request.cache_id} "
+                        f"session-entry mode={plan.mode} lcp={plan.lcp} "
+                        f"cached={plan.cached_len} -> prefix={pid} reused={fallback.reused}",
+                        flush=True,
+                    )
+                    # Same divergence preview the full-prefill path prints:
+                    # what the session entry and this prompt disagree on is
+                    # the only thing that says which side to fix.
+                    try:
+                        _tk = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+                        _cached = list(getattr(sub.seed_state, "token_ids", None) or [])
+                        _lo = max(0, plan.lcp - 6)
+                        print(
+                            f"[batch] churn cache_id={sub.request.cache_id} at={plan.lcp} "
+                            f"cached={_tk.decode(_cached[_lo:plan.lcp + 24])!r} "
+                            f"prompt={_tk.decode(list(sub.prompt_tokens)[_lo:plan.lcp + 24])!r}",
+                            flush=True,
+                        )
+                    except Exception as exc:  # noqa: BLE001 — diagnostics never break a turn
+                        print(f"[batch] churn preview unavailable: {exc}", flush=True)
+                    sub.seed_state = seeded
+                    plan = fallback
+                    break
+        sub.prefill_total = len(plan.segment)
+        sub.reused_tokens = plan.reused
+        sub.seed_mode = plan.mode
+        # Publish the shared band ONLY from a session that reused nothing —
+        # the "pioneer" of this prefix; every sibling after it inherits the
+        # band and skips it. The band is a SECOND cut beside the session's
+        # own end-minus-margin boundary, not a replacement for it: saving the
+        # pioneer at the band made its next turn re-prefill everything after
+        # it, which on Gemma (tool block rendered after the system text) was
+        # 7,313 of 8,604 tokens on the second message of every chat.
+        if sub.snapshot_target is None and str(plan.mode).startswith("fresh"):
+            sub.band_target = _band_snapshot_target(sub)
+        if sub.request.cache_id:
+            print(
+                f"[batch] seed cache_id={sub.request.cache_id} mode={plan.mode} "
+                f"reused={plan.reused} prefill={len(plan.segment)}",
+                flush=True,
+            )
+            if plan.mode == "fresh-untrimmable" and plan.cached_len:
+                # Name the churn instead of leaving it to inference. Two
+                # rounds of this investigation guessed the cause from cache
+                # SIZES (which say what was available, not what matched) and
+                # both guesses were wrong; decoding the two sides at the
+                # divergence point is the only thing that identifies it.
+                try:
+                    _tk = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+                    _cached = list(getattr(sub.seed_state, "token_ids", None) or [])
+                    _lo = max(0, plan.lcp - 6)
+                    print(
+                        f"[batch] churn cache_id={sub.request.cache_id} "
+                        f"at={plan.lcp} "
+                        f"cached={_tk.decode(_cached[_lo:plan.lcp + 24])!r} "
+                        f"prompt={_tk.decode(list(sub.prompt_tokens)[_lo:plan.lcp + 24])!r}",
+                        flush=True,
+                    )
+                except Exception as exc:  # noqa: BLE001 — diagnostics never break a turn
+                    print(f"[batch] churn preview unavailable: {exc}", flush=True)
+            if plan.mode == "fresh-untrimmable":
+                # lcp/cached is the diagnostic pair: lcp near 0 means the
+                # prompt churned early (fix the prompt), lcp near cached
+                # means the snapshot boundary was cut past the stable
+                # region (fix the boundary). Same symptom, opposite fixes.
+                print(
+                    f"[batch] WARNING full prefill: cache_id={sub.request.cache_id} "
+                    f"had divergent untrimmable state without a reusable prompt "
+                    f"snapshot (tokens={len(plan.segment)} lcp={plan.lcp} "
+                    f"cached={plan.cached_len} "
+                    f"lcp_frac={plan.lcp / plan.cached_len:.2f})"
+                    if plan.cached_len
+                    else f"[batch] WARNING full prefill: cache_id={sub.request.cache_id} "
+                    f"had divergent untrimmable state (tokens={len(plan.segment)})",
+                    flush=True,
+                )
+                print(
+                    f"[batch] churn-shape cache_id={sub.request.cache_id} "
+                    f"probe_roster={sub.roster_fp} "
+                    f"saved_roster={_LAST_SAVE_ROSTER.get(sub.request.cache_id, 'unknown')} "
+                    f"path={'warm' if str(sub.request_id).startswith('cachewarm-') else 'chat'} "
+                    f"prompt_len={len(sub.prompt_tokens)}",
+                    flush=True,
+                )
+        return plan.segment, plan.caches, plan.all_tokens
+
+    def _snapshot_segments(self, sub, segment):
+        """Plant the per-sequence segment edges for KV extraction: the
+        session's own boundary, plus the shared band's on a pioneer turn.
+
+        BatchGenerator reports every prefill chunk as a prompt response. The
+        old handler copied the growing cache at *every* chunk and kept only the
+        last copy. Long prompts therefore paid dozens of large, discarded KV
+        copies. Dedicated targets make capture one-shot per boundary and also
+        work for unequal-length multi-sequence waves: mlx_lm finalizes each
+        chunk's right-padding before returning, then extract_cache(uid)
+        removes that sequence's left-padding.
+        """
+        if not self._needs_snapshot or not sub.request.cache_id:
+            sub.band_target = None
+            return [segment]
+        plan = cache_seed.plan_snapshot_segments(
+            segment,
+            sub.reused_tokens,
+            len(sub.prompt_tokens),
+            sub.snapshot_target,
+            _SNAPSHOT_BOUNDARY_MARGIN,
+            sub.band_target,
+        )
+        sub.snapshot_target = plan.target
+        sub.band_target = plan.band
+        return plan.segments
+
+    def _emit_prefill_liveness(self) -> None:
+        """While the active wave is still prefilling (no token emitted yet),
+        print a `Prefill:` line the supervisor's stdout-parser turns into a
+        `prefill` phase event — keeping the TS pre-first-byte watchdog anchored
+        on real engine liveness instead of tripping its silent-cold-start
+        bound. Honest by construction: it only fires while `_prefill_total > 0`
+        (cleared the instant the first token lands), and only from the live
+        worker loop — a genuinely wedged engine emits nothing and still trips.
+
+        Progress is real: the worker consumes `BatchGenerator.next()` per
+        internal step, so each prompt chunk reports its (processed, total)
+        and the marker carries an actual percentage instead of a constant
+        0%.
+
+        Each marker names the sub it belongs to (`cache=<cache_id>`), and
+        carries THAT sub's own done/total rather than the wave aggregate —
+        see `_prefill_meta`. An untagged marker still parses, and the
+        supervisor still broadcasts it, so an engine without the tag degrades
+        to the old behaviour rather than going dark."""
+        total = self._prefill_total
+        if total < _PREFILL_LIVENESS_MIN_TOKENS:
+            return
+        now = time.time()
+        if now - self._last_prefill_emit < _PREFILL_LIVENESS_INTERVAL_S:
+            return
+        self._last_prefill_emit = now
+        # Format MUST match stdout-parser's Prefill regex:
+        #   Prefill:  42%|   | 8192/<total> [batched] cache=<id>
+        if not self._prefill_meta:
+            done = min(total, sum(self._prefill_done.values()))
+            pct = min(99, (100 * done) // total) if total > 0 else 0
+            print(f"Prefill: {pct:3d}%|          | {done}/{total} [batched]", flush=True)
+            return
+        for key, (cache_id, sub_total) in self._prefill_meta.items():
+            sub_total = max(0, int(sub_total or 0))
+            if sub_total <= 0:
+                continue
+            done = min(sub_total, int(self._prefill_done.get(key, 0)))
+            pct = min(99, (100 * done) // sub_total)
+            tag = f" cache={cache_id}" if cache_id else ""
+            print(
+                f"Prefill: {pct:3d}%|          | {done}/{sub_total} [batched]{tag}",
+                flush=True,
+            )
+
+    def _admit_count(self, pending: int) -> int:
+        """How many of the `pending` co-arrived subs to start THIS wave.
+
+        `self._max` is the hard concurrency contract advertised by the engine;
+        pending work beyond it must remain queued for the next static wave.
+        Within that width, live allocation is a second backstop to the TS-side
+        slot estimate: when the device is already tight, collapse the wave to
+        one request rather than risking a Metal abort."""
+        capped = min(pending, self._max)
+        if capped <= 1 or _MEM_LIMIT_BYTES <= 0:
+            return capped
+        threshold = _MEM_LIMIT_BYTES * _MEM_ADMIT_FRACTION
+        active = _safe_active_memory()
+        if active <= 0:
+            return capped  # can't read usage → keep the configured hard cap
+        if active >= threshold:
+            # Already tight before this wave even starts — reclaim any freed
+            # buffer cache and re-check once before forcing serial.
+            _reclaim_mlx_buffer_cache("batch-admission", force=True)
+            active = _safe_active_memory()
+        return 1 if active >= threshold else capped
+
+    async def _run(self):
+        while True:
+            if not self._pending and not self._subs:
+                self._wake.clear()
+                await self._wake.wait()
+            async with _get_generation_lock():
+                await self._step_wave()
+                while self._subs:
+                    await self._step_wave()
+
+    async def _step_wave(self):
+        # Evict client-disconnected sequences before stepping.
+        gone = [uid for uid, s in self._subs.items() if s.cancelled]
+        if gone:
+            for uid in gone:
+                self._save_cancelled_snapshot(self._subs[uid])
+            try:
+                self._gen.remove(gone)
+            except Exception as exc:  # noqa: BLE001
+                print(f"[batch] remove failed: {exc}", flush=True)
+            for uid in gone:
+                self._subs.pop(uid, None)
+
+        # Static-wave admission: only start a new batch when the current
+        # one has fully drained. mlx_lm's continuous merge of a mid-flight
+        # prompt batch into the running generation batch isn't supported by
+        # the mlx_vlm model caches here (raises broadcast (N),(M)), so we
+        # batch co-arriving requests into a wave and run it to completion
+        # before admitting the next. Co-arriving turns (multi-pane,
+        # ask-chains) still share batched decode; a turn that arrives
+        # mid-wave waits for the wave to drain — no worse than the old
+        # serial path, and strictly better when turns arrive together.
+        if self._pending and not self._subs:
+            self._drop_cancelled_pending()
+            if not self._pending:
+                return
+            # Interactive requests first, and never in a wave with background
+            # work — see wave_policy.select_wave.
+            batch, self._pending = wave_policy.select_wave(self._pending, self._admit_count)
+            if self._pending:
+                # Deferred by the configured concurrency ceiling, live
+                # memory pressure, or because the wave is interactive-only.
+                # The remaining subs run in the next wave, after this one
+                # drains and frees its KV. The outer loop re-enters admission
+                # once _subs empties (it never blocks on _wake while _pending
+                # is non-empty).
+                held_bg = sum(1 for s in self._pending if wave_policy.is_background(s))
+                print(
+                    f"[batch] admission-throttled: admitting {len(batch)}, "
+                    f"deferring {len(self._pending)} to next wave "
+                    f"({held_bg} of them background, "
+                    f"max_concurrency={self._max}, "
+                    f"active={_safe_active_memory() // (1024 * 1024)}MB, "
+                    f"ceiling={_MEM_LIMIT_BYTES // (1024 * 1024)}MB)",
+                    flush=True,
+                )
+            for sub in batch:
+                self._announce_admitted(sub)
+            if len(batch) == 1 and _SPEC is not None and self._drafter_busy:
+                print(
+                    f"[spec] off request={batch[0].request_id} reason=the drafter "
+                    f"belongs to a paused wave",
+                    flush=True,
+                )
+            elif len(batch) == 1 and _SPEC is not None:
+                spec_sub = batch[0]
+                mode, why = spec_decode.spec_mode(
+                    spec_sub.request, spec_sub.grammar
+                )
+                if mode is not None:
+                    admit, gate_why = spec_decode.draft_prefill_gate(
+                        _SPEC, len(spec_sub.prompt_tokens)
+                    )
+                    if not admit:
+                        mode, why = None, gate_why
+                if mode is not None:
+                    await self._run_spec_wave(spec_sub, mode)
+                    return
+                print(
+                    f"[spec] off request={spec_sub.request_id} reason={why}",
+                    flush=True,
+                )
+            segments, caches, alltoks, samplers, lps, maxtoks = [], [], [], [], [], []
+            for sub in batch:
+                seg, c, at = self._seed_args(sub)
+                segments.append(self._snapshot_segments(sub, seg))
+                caches.append(c)
+                alltoks.append(at)
+                samplers.append(self._sampler_for(sub.request))
+                lps.append(self._processors_for(sub.request, sub.grammar))
+                maxtoks.append(int(sub.request.max_tokens) if sub.request.max_tokens else 2048)
+            try:
+                uids = self._gen.insert_segments(
+                    segments=segments,
+                    max_tokens=maxtoks,
+                    caches=caches,
+                    all_tokens=alltoks,
+                    samplers=samplers,
+                    logits_processors=lps,
+                )
+            except Exception as exc:  # noqa: BLE001
+                print(f"[batch] insert failed: {exc}", flush=True)
+                log_contained_exception("batch")
+                for sub in batch:
+                    sub.queue.put_nowait(("err", str(exc)))
+                return
+            for uid, sub in zip(uids, batch):
+                sub.uid = uid
+                sub.prefill_started_at = time.perf_counter()
+                self._subs[uid] = sub
+            # Arm prefill-liveness for the wave: total new tokens to
+            # prefill across its subs (authoritative — from the seed
+            # plans above). Reset the throttle so the first marker
+            # emits immediately (before the blocking decode below).
+            self._prefill_total = sum(
+                max(0, int(getattr(s, "prefill_total", 0))) for s in batch
+            )
+            self._prefill_done = {}
+            self._prefill_meta = {
+                uid: (
+                    getattr(s.request, "cache_id", None),
+                    max(0, int(getattr(s, "prefill_total", 0))),
+                )
+                for uid, s in self._subs.items()
+            }
+            self._last_prefill_emit = 0.0
+
+        # One engine step across all active sequences. `next()` (not
+        # `next_generated()`) so control returns after EVERY internal
+        # step — prompt chunks included. That is what makes the
+        # end-of-prompt snapshot capturable at its exact boundary
+        # (see _capture_prompt_snapshot), gives the liveness marker
+        # real per-chunk progress instead of a constant 0%, and lets
+        # the event loop breathe between prefill chunks.
+        # A wave of background work steps aside, between steps, for a person
+        # who has been waiting since it started. It resumes from this exact
+        # step once they are served.
+        if self._subs and self._preempt_due(self._subs.values()):
+            await self._serve_preemptors(list(self._subs.values()))
+        if self._subs:
+            # Emit a prefill marker before the (potentially long, event-
+            # loop-blocking) step so the watchdog sees liveness.
+            # No-op once the first token lands (_prefill_total → 0).
+            self._emit_prefill_liveness()
+            self._emit_waiting_liveness()
+            try:
+                prompt_responses, responses = self._gen.next()
+                step_ended_at = time.perf_counter()
+            except Exception as exc:  # noqa: BLE001
+                print(f"[batch] step failed: {exc}", flush=True)
+                log_contained_exception("batch")
+                for sub in list(self._subs.values()):
+                    sub.queue.put_nowait(("err", str(exc)))
+                self._subs.clear()
+                return
+            # set_cache_limit bounds future reuse, while this periodic
+            # pressure check also catches temporary overshoot during a
+            # long prefill/decode. It replaces reliance on mlx_lm's much
+            # coarser 512-generation-step cache clear.
+            self._memory_check_steps += 1
+            if self._memory_check_steps >= 32:
+                self._memory_check_steps = 0
+                _reclaim_mlx_buffer_cache("batch-pressure")
+            for pr in prompt_responses:
+                progress = getattr(pr, "progress", None)
+                has_progress = isinstance(progress, tuple) and len(progress) == 2
+                if has_progress:
+                    self._prefill_done[pr.uid] = int(progress[0])
+                psub = self._subs.get(pr.uid)
+                if psub is None or not self._needs_snapshot or not psub.request.cache_id:
+                    continue
+                if getattr(pr, "end_of_prompt", False):
+                    continue
+                if has_progress and getattr(pr, "end_of_segment", False):
+                    # Capture only the deliberately planted edge. Chunk
+                    # edges are observable too, but copying at each one is
+                    # expensive and unnecessary. extract_cache(uid) is
+                    # padding-aware after PromptProcessingBatch.prompt()
+                    # finalizes this step, so every sub in a multi-request
+                    # wave can safely take its own snapshot.
+                    boundary = psub.reused_tokens + int(progress[0])
+                    if boundary in (psub.snapshot_target, psub.band_target):
+                        self._capture_prompt_snapshot(psub, boundary)
+            if responses:
+                # A token came back → prefill for this wave is done; stop
+                # the liveness markers (the SSE first byte now drives the
+                # watchdog's tighter streaming-idle bound).
+                self._prefill_total = 0
+                self._prefill_done = {}
+                self._prefill_meta = {}
+            for r in responses:
+                sub = self._subs.get(r.uid)
+                if sub is None:
+                    continue
+                if sub.generation_started_at is None:
+                    # The step that yields the first token is the closest
+                    # boundary BatchGenerator exposes between prefill and
+                    # decode. Count it with prefill/TTFT, then measure decode
+                    # from the inter-token intervals that follow; otherwise
+                    # a short prompt can produce a near-zero denominator and
+                    # a wildly inflated first-token rate.
+                    sub.generation_started_at = step_ended_at
+                    prefill_started_at = sub.prefill_started_at or step_ended_at
+                    prefill_seconds = max(
+                        step_ended_at - prefill_started_at,
+                        1e-9,
+                    )
+                    if sub.prefill_total > 0:
+                        sub.prompt_tps = sub.prefill_total / prefill_seconds
+                sub.first_token_seen = True
+                sub.token_ids.append(int(r.token))
+                generated_intervals = len(sub.token_ids) - 1
+                generation_seconds = step_ended_at - sub.generation_started_at
+                if generated_intervals > 0 and generation_seconds > 0:
+                    sub.generation_tps = generated_intervals / generation_seconds
+                self._emit_delta(sub)
+                if r.finish_reason is not None:
+                    self._finish(sub, r)
+                    self._subs.pop(r.uid, None)
+            # Yield so SSE generators flush + new requests get admitted.
+            await self._breathe()
+
+
+    async def _run_spec_wave(self, sub, mode="greedy"):
+        """Serve one eligible sub via MTP speculation (sidecar-owned wave).
+
+        Replaces BatchGenerator for this wave only: seed plan + direct
+        chunked prefill (with the planted boundary snapshot) + upstream's
+        serial _mtp_rounds. Emit, liveness, cancellation, stats, and the
+        finish-time save reuse the engine's existing machinery, so a spec
+        turn is indistinguishable downstream from a batch turn. Measured
+        basis + eligibility rationale live in spec_decode.py."""
+        sub.prefill_started_at = time.perf_counter()
+        self._spec_sub = sub
+        try:
+            # SeedPlan's third field is the tokens ALREADY IN the seeded
+            # cache (the reused prefix — empty on fresh), mirroring what
+            # insert_segments' all_tokens means to mlx_lm.
+            seg, seeded, prefix_toks = self._seed_args(sub)
+            tower = getattr(MODEL, "language_model", MODEL)
+            cache_layers = seeded
+            if cache_layers is not None and not spec_decode.cache_matches_tower(
+                cache_layers, tower
+            ):
+                print(
+                    f"[spec] seed-guard: cached layer classes mismatch the "
+                    f"active tower for cache_id={sub.request.cache_id}; "
+                    f"treating as fresh",
+                    flush=True,
+                )
+                cache_layers = None
+                seg = list(sub.prompt_tokens)
+                prefix_toks = []
+                sub.reused_tokens = 0
+                sub.prefill_total = len(seg)
+                sub.seed_mode = "fresh"
+            if cache_layers is None:
+                cache_layers = tower.make_cache()
+            full_prompt = list(prefix_toks or []) + list(seg)
+            max_toks = (
+                int(sub.request.max_tokens) if sub.request.max_tokens else 2048
+            )
+
+            cuts: List[int] = []
+            if self._needs_snapshot and sub.request.cache_id:
+                plan = cache_seed.plan_snapshot_segments(
+                    seg,
+                    sub.reused_tokens,
+                    len(sub.prompt_tokens),
+                    sub.snapshot_target,
+                    _SNAPSHOT_BOUNDARY_MARGIN,
+                    sub.band_target,
+                )
+                sub.snapshot_target = plan.target
+                sub.band_target = plan.band
+                edge = 0
+                for part in plan.segments[:-1]:
+                    edge += len(part)
+                    cuts.append(edge)
+            else:
+                sub.band_target = None
+
+            self._prefill_total = max(0, int(sub.prefill_total or 0))
+            self._prefill_done = {}
+            self._prefill_meta = {
+                0: (
+                    getattr(sub.request, "cache_id", None),
+                    max(0, int(sub.prefill_total or 0)),
+                )
+            }
+            self._last_prefill_emit = 0.0
+
+            final_out = None
+            for kind, done, _total, maybe_out in spec_decode.chunked_prefill_steps(
+                tower, cache_layers, list(seg), ARGS.prefill_step_size, cuts=cuts
+            ):
+                if sub.cancelled:
+                    print(
+                        f"[spec] wave cancelled during prefill "
+                        f"request={sub.request_id}",
+                        flush=True,
+                    )
+                    self._save_cancelled_snapshot(sub, cache_layers, full_prompt, done)
+                    return
+                self._prefill_done[0] = int(done)
+                if kind == "cut":
+                    self._capture_spec_snapshot(
+                        sub, cache_layers, full_prompt, int(sub.reused_tokens) + int(done)
+                    )
+                elif kind == "final":
+                    final_out = maybe_out
+                # Chunk edges are quiescent: the cache holds exactly the
+                # tokens prefilled so far, and the drafter is not in use yet.
+                if self._spec_preempt_due(sub):
+                    await self._serve_preemptors([sub])
+                self._emit_prefill_liveness()
+                self._emit_waiting_liveness()
+                await self._breathe()
+
+            self._prefill_total = 0
+            self._prefill_done = {}
+            self._prefill_meta = {}
+
+            # Assisted mode: positioned target sampling (temperature/top_p/
+            # top_k, per-(seed,position) keys) and/or the processed walk.
+            # The processor list is built by the SAME _processors_for the
+            # normal path uses, so constrained decoding is identical either
+            # route. Seed is logged for replay.
+            sampler = None
+            processors = None
+            history = None
+            wave_seed = None
+            if mode == "assisted":
+                processors = self._processors_for(sub.request, sub.grammar) or None
+                t = sub.request.temperature
+                if t not in (None, 0, 0.0):
+                    wave_seed = int.from_bytes(os.urandom(6), "big")
+                    sampler = spec_decode.PositionedSampler(
+                        temperature=float(t),
+                        top_p=float(sub.request.top_p)
+                        if sub.request.top_p is not None
+                        else 0.0,
+                        top_k=int(sub.request.top_k)
+                        if sub.request.top_k is not None
+                        else 0,
+                        seed=wave_seed,
+                    )
+                history = list(full_prompt)
+
+            def emit(tok):
+                now = time.perf_counter()
+                if sub.generation_started_at is None:
+                    # Mirror the batch worker: the step yielding the first
+                    # token counts with prefill/TTFT; decode is measured
+                    # from inter-token intervals after it.
+                    sub.generation_started_at = now
+                    prefill_seconds = max(
+                        now - (sub.prefill_started_at or now), 1e-9
+                    )
+                    if sub.prefill_total > 0:
+                        sub.prompt_tps = sub.prefill_total / prefill_seconds
+                sub.first_token_seen = True
+                sub.token_ids.append(int(tok))
+                intervals = len(sub.token_ids) - 1
+                gen_seconds = now - sub.generation_started_at
+                if intervals > 0 and gen_seconds > 0:
+                    sub.generation_tps = intervals / gen_seconds
+                self._emit_delta(sub)
+
+            finish_reason = None
+            if mode == "assisted":
+                first = spec_decode.processed_first_token(
+                    final_out, sampler, processors, history
+                )
+                history.append(first)
+            else:
+                first = spec_decode.first_token_from(final_out)
+            emit(first)
+            if first in self._stop_ids:
+                finish_reason = "stop"
+            elif len(sub.token_ids) >= max_toks:
+                finish_reason = "length"
+
+            if finish_reason is None:
+                # From here the drafter's per-wave state (its cache, shared
+                # kv, position) belongs to this wave until it ends.
+                self._drafter_busy = True
+                if mode == "assisted":
+                    rounds = spec_decode.assisted_rounds(
+                        MODEL,
+                        _SPEC,
+                        cache_layers,
+                        list(seg),
+                        final_out,
+                        first,
+                        max_toks,
+                        sampler=sampler,
+                        processors=processors,
+                        history=history,
+                    )
+                else:
+                    rounds = spec_decode.make_rounds(
+                        MODEL,
+                        _SPEC,
+                        cache_layers,
+                        list(seg),
+                        final_out,
+                        first,
+                        max_toks,
+                    )
+                mem_steps = 0
+                for tok, _lp in rounds:
+                    if sub.cancelled:
+                        print(
+                            f"[spec] wave cancelled request={sub.request_id}",
+                            flush=True,
+                        )
+                        self._save_cancelled_snapshot(sub)
+                        return
+                    emit(tok)
+                    if tok in self._stop_ids:
+                        finish_reason = "stop"
+                        break
+                    if len(sub.token_ids) >= max_toks:
+                        finish_reason = "length"
+                        break
+                    mem_steps += 1
+                    if mem_steps >= 32:
+                        mem_steps = 0
+                        _reclaim_mlx_buffer_cache("spec-pressure")
+                    # Each yield follows a committed verify round (upstream
+                    # commits before emitting), so the target cache is
+                    # consistent here. A wave run inside the pause cannot
+                    # draft: _drafter_busy keeps it on the BatchGenerator.
+                    if self._spec_preempt_due(sub):
+                        await self._serve_preemptors([sub])
+                    self._emit_waiting_liveness()
+                    await self._breathe()
+                if finish_reason is None:
+                    finish_reason = (
+                        "length" if len(sub.token_ids) >= max_toks else "stop"
+                    )
+
+            line = spec_decode.stats_line(_SPEC)
+            if line:
+                print(
+                    f"[spec] stats request={sub.request_id} mode={mode} "
+                    + (f"seed={wave_seed} " if wave_seed is not None else "")
+                    + f"ctx={len(sub.prompt_tokens)} gen={len(sub.token_ids)} "
+                    f":: {line}",
+                    flush=True,
+                )
+            # Rounds can exit mid-round (stop token, budget) with verified-
+            # but-rejected tokens still in the KV. Hybrid stacks are immune —
+            # their save always uses the boundary snapshot — but a trimmable
+            # stack must be trimmed back to exactly the tokens we report, or
+            # the saved entry's offsets outrun its token list.
+            try:
+                reported = len(sub.prompt_tokens) + len(sub.token_ids)
+                offsets = [
+                    int(getattr(c, "offset", 0) or 0) for c in cache_layers
+                ]
+                excess = max(offsets) - reported if offsets else 0
+                if excess > 0:
+                    cache_seed.trim_layers(cache_layers, excess)
+            except Exception:  # noqa: BLE001 — snapshot path covers hybrids
+                pass
+            r = types.SimpleNamespace(
+                finish_reason=finish_reason,
+                prompt_cache=list(cache_layers),
+                all_tokens=full_prompt + [int(t) for t in sub.token_ids],
+                token=sub.token_ids[-1] if sub.token_ids else 0,
+            )
+            self._finish(sub, r)
+        except Exception as exc:  # noqa: BLE001
+            print(f"[spec] wave failed request={sub.request_id}: {exc}", flush=True)
+            # Tag every traceback line. The supervisor's stdout classifier
+            # reads a bare `SomeError: ...` leaf line as proof the engine is
+            # dead and tears the process down — right during startup, wrong
+            # here, where this handler has already contained the failure and
+            # the engine is still serving other requests.
+            log_contained_exception("spec")
+            sub.queue.put_nowait(("err", str(exc)))
+        finally:
+            self._drafter_busy = False
+            if self._spec_sub is sub:
+                self._spec_sub = None
+
+    def _spec_preempt_due(self, sub):
+        if self._nested or not wave_policy.is_background(sub):
+            return False
+        return self._preempt_due([sub])
+
+    def _capture_spec_snapshot(self, sub, cache_layers, full_prompt, boundary):
+        """Boundary snapshot for the spec wave's direct prefill. The cache
+        is quiescent at the planted chunk edge; clone_layers reconstructs
+        standalone layers (MLX arrays are immutable, so sharing buffers is
+        safe), and the same token-prefix proof as _capture_prompt_snapshot
+        guards the boundary invariant."""
+        try:
+            boundary = int(boundary)
+            layers = spec_decode.clone_layers(cache_layers)
+            if cache_seed.snapshot_matches_prompt(
+                layers, full_prompt[:boundary], sub.prompt_tokens, boundary
+            ):
+                _store_snapshot(sub, (layers, boundary))
+            elif not self._snapshot_warned:
+                self._snapshot_warned = True
+                print(
+                    f"[spec] prompt snapshot skipped: boundary={boundary} "
+                    f"(post-gen save fallback)",
+                    flush=True,
+                )
+        except Exception as exc:  # noqa: BLE001 — snapshot is best-effort
+            if not self._snapshot_warned:
+                self._snapshot_warned = True
+                print(f"[spec] prompt snapshot failed: {exc}", flush=True)
+
+    def _capture_prompt_snapshot(self, sub, at_tokens):
+        """Capture the KV state at a mid-prompt boundary of `at_tokens`.
+
+        Called from the worker's prompt-response handling at segment /
+        chunk edges (the worker consumes `BatchGenerator.next()` per
+        internal step, so the batch is quiescent between calls). A
+        segment edge is a token-index cut deep inside text both turns
+        share, so it is immune to the BPE boundary problem — measured
+        live: the end-of-prompt token sequence `…<start_of_turn>model\\n`
+        re-merges differently once the next turn's text continues past
+        it, so a full-length snapshot misses by ~2 tokens and wrapped
+        rotating layers can't trim even that. The admission loop plants
+        a deliberate cut `_SNAPSHOT_BOUNDARY_MARGIN` tokens before the
+        prompt's end so every turn gets one stable, advancing boundary.
+
+        `extract_cache` materializes contiguous per-sequence copies and
+        returns that sequence's unpadded token list, so the snapshot doesn't
+        alias the live batch buffers or accidentally retain another sub's
+        right-padding. Exact token-prefix + layer-offset checks are the runtime
+        proof of the boundary invariant — on any mismatch the snapshot is
+        dropped (logged once) and _finish falls back to the post-gen save.
+        The same edge may be the session boundary, the band boundary, or both;
+        `_store_snapshot` files it under whichever it is.
+        """
+        try:
+            got = self._gen.extract_cache([sub.uid]).get(sub.uid)
+            layers = got[0] if got else None
+            extracted_tokens = got[1] if got else None
+            if cache_seed.snapshot_matches_prompt(
+                layers, extracted_tokens, sub.prompt_tokens, at_tokens
+            ):
+                _store_snapshot(sub, (list(layers), at_tokens))
+            elif not self._snapshot_warned:
+                self._snapshot_warned = True
+                offsets = [getattr(c, "offset", None) for c in (layers or [])]
+                print(
+                    f"[batch] prompt snapshot skipped: boundary={at_tokens} "
+                    f"extracted_tokens={len(extracted_tokens or [])} "
+                    f"offsets={offsets} (post-gen save fallback)",
+                    flush=True,
+                )
+        except Exception as exc:  # noqa: BLE001 — snapshot is best-effort
+            if not self._snapshot_warned:
+                self._snapshot_warned = True
+                print(f"[batch] prompt snapshot failed: {exc}", flush=True)
+
+    def _emit_delta(self, sub):
+        # Full-decode-delta detokenization: correct across BPE merges /
+        # multibyte chars (hold on a trailing replacement char until the
+        # next token completes it). Fine for chat-length outputs.
+        full = self._tok.decode(sub.token_ids)
+        if full.endswith("�"):
+            return
+        # Strip leaked framing tokens before computing the delta. The scrub
+        # is deterministic over the cumulative decode, so tracking `emitted`
+        # against the scrubbed string stays consistent across steps.
+        full = _scrub_leaked_markers(full)
+        delta = full[sub.emitted :]
+        if not delta:
+            return
+        sub.emitted = len(full)
+        sub.queue.put_nowait(
+            (
+                "tok",
+                delta,
+                len(sub.token_ids),
+                sub.prompt_tps,
+                sub.generation_tps,
+            )
+        )
+
+    def _finish(self, sub, r):
+        full = _scrub_leaked_markers(self._tok.decode(sub.token_ids))
+        if len(full) > sub.emitted:
+            sub.queue.put_nowait(
+                (
+                    "tok",
+                    full[sub.emitted :],
+                    len(sub.token_ids),
+                    sub.prompt_tps,
+                    sub.generation_tps,
+                )
+            )
+            sub.emitted = len(full)
+        try:
+            if sub.request.cache_id:
+                # On finish the sequence is already out of the generation
+                # batch, so extract_cache(uid) misses — the Response carries
+                # the final KV (`prompt_cache`) + full token list
+                # (`all_tokens`). Fall back to a token reconstruction.
+                cache_layers = getattr(r, "prompt_cache", None)
+                all_tokens = getattr(r, "all_tokens", None)
+                if cache_layers is not None:
+                    post_gen = list(cache_layers)
+                    # Save decision: a post-generation cache is only worth
+                    # keeping if a future turn can trim it back to wherever
+                    # the re-templated history diverges from these raw
+                    # tokens. Un-trimmable stacks (wrapped windowed layers)
+                    # get the mid-prompt snapshot instead — a stable prefix
+                    # of the next prompt, no trim ever needed. A tiny turn
+                    # whose segment was too short to capture one keeps the
+                    # PRIOR entry (still a valid prefix) rather than
+                    # overwriting it with un-trimmable full-length state.
+                    trimmable = cache_seed.all_trimmable(post_gen)
+                    if not trimmable and sub.prompt_snapshot is None and sub.reused_tokens > 0:
+                        prior_entry = _CACHE.get(sub.request.cache_id)
+                        if prior_entry is not None:
+                            sub.saved_cache_state = prior_entry.state
+                        print(
+                            f"[batch] kept prior cache entry for "
+                            f"cache_id={sub.request.cache_id} (segment too short "
+                            f"for a stable snapshot; post-turn cache not trimmable)",
+                            flush=True,
+                        )
+                    else:
+                        use_snapshot = sub.prompt_snapshot is not None and not trimmable
+                        state = PromptCacheState()
+                        if use_snapshot:
+                            snap_layers, snap_count = sub.prompt_snapshot
+                            state.cache = snap_layers
+                            state.token_ids = list(sub.prompt_tokens[:snap_count])
+                        else:
+                            state.cache = post_gen
+                            state.token_ids = (
+                                [int(t) for t in all_tokens]
+                                if all_tokens is not None
+                                else list(sub.prompt_tokens) + list(sub.token_ids)
+                            )
+                        if state.token_ids:
+                            _save_cache(sub.request.cache_id, state)
+                            sub.saved_cache_state = state
+                            _LAST_SAVE_ROSTER[sub.request.cache_id] = sub.roster_fp
+                            prefix_state = _prefix_seed_state(sub, state, trimmable)
+                            if prefix_state is not None:
+                                _seed_prefix_from_session(sub.request, prefix_state)
+                            print(
+                                f"[batch] saved cache_id={sub.request.cache_id} "
+                                f"tokens={len(state.token_ids)}"
+                                + (
+                                    " (prompt snapshot; post-turn cache not trimmable)"
+                                    if use_snapshot
+                                    else ""
+                                ),
+                                flush=True,
+                            )
+                # Reuse parity with the serial path's terminal log — the
+                # line that makes "is the cache actually working?" a one-
+                # line grep instead of a timing investigation.
+                print(
+                    f"[cache] reuse cache_id={sub.request.cache_id} "
+                    f"cached_tokens={sub.reused_tokens} "
+                    f"prefilled_tokens={sub.prefill_total} [batched]",
+                    flush=True,
+                )
+        except Exception as exc:  # noqa: BLE001
+            print(f"[batch] cache save failed: {exc}", flush=True)
+        sub.queue.put_nowait(
+            ("done", r.finish_reason, sub.prompt_tps, sub.generation_tps)
+        )
+
+
+def _batched_tool_payload(frag, rid: str, model: str, in_toks: int, out_count: int, sub) -> Dict[str, Any]:
+    """Batched-route envelope for one tool-stream fragment.
+
+    Mirrors `_tool_stream_payload` but sources usage from the sub rather than
+    a generation chunk — the batched worker reports counters separately.
+    """
+    if frag.content is not None:
+        delta: Dict[str, Any] = {"role": "assistant", "content": frag.content}
+    else:
+        call: Dict[str, Any] = {
+            "index": frag.tool_call_index,
+            "function": {"arguments": frag.arguments or ""},
+        }
+        if frag.tool_call_id is not None:
+            call["id"] = frag.tool_call_id
+            call["type"] = "function"
+            call["function"]["name"] = frag.name or ""
+        delta = {"role": "assistant", "tool_calls": [call]}
+    return {
+        "id": rid,
+        "object": "chat.completion.chunk",
+        "created": int(time.time()),
+        "model": model,
+        "choices": [{"index": 0, "delta": delta, "finish_reason": None}],
+        "usage": {
+            "input_tokens": in_toks,
+            "output_tokens": out_count,
+            "total_tokens": in_toks + out_count,
+            "prompt_tps": 0.0,
+            "generation_tps": 0.0,
+            "cached_tokens": sub.reused_tokens,
+        },
+    }
+
+
+async def _batched_stream_iter(sub: "_Sub"):
+    """SSE generator for the batched path — same wire frames as the serial
+    route, fed from the engine worker's per-sub queue."""
+    request = sub.request
+    rid = sub.request_id
+    in_toks = len(sub.prompt_tokens)
+    # Same structured tool-call split as the serial route. The batched path
+    # is the one the eval harness actually exercises (max-concurrency 1 still
+    # runs through the batch engine), so wiring only the serial route left
+    # the flag inert.
+    tool_stream = _make_tool_stream(rid, getattr(request, 'tools', None))
+    try:
+        while True:
+            item = await sub.queue.get()
+            kind = item[0]
+            if kind == "tok":
+                _, delta, out_count, prompt_tps, generation_tps = item
+                if await sub.http_request.is_disconnected():
+                    sub.cancelled = True
+                    return
+                if tool_stream is not None:
+                    for _frag in tool_stream.feed(delta):
+                        yield "data: " + json.dumps(
+                            _batched_tool_payload(_frag, rid, request.model, in_toks, out_count, sub)
+                        ) + "\n\n"
+                    continue
+                yield "data: " + json.dumps(
+                    {
+                        "id": rid,
+                        "object": "chat.completion.chunk",
+                        "created": int(time.time()),
+                        "model": request.model,
+                        "choices": [
+                            {
+                                "index": 0,
+                                "delta": {"role": "assistant", "content": delta},
+                                "finish_reason": None,
+                            }
+                        ],
+                        "usage": {
+                            "input_tokens": in_toks,
+                            "output_tokens": out_count,
+                            "total_tokens": in_toks + out_count,
+                            "prompt_tps": prompt_tps,
+                            "generation_tps": generation_tps,
+                            # Cache tokens actually served per the seed
+                            # plan — parity with the serial path so the
+                            # TS session's cachedInputTokens works on
+                            # both. 0 means a genuinely cold prefill.
+                            "cached_tokens": sub.reused_tokens,
+                        },
+                    }
+                ) + "\n\n"
+            elif kind == "done":
+                _, reason, prompt_tps, generation_tps = item
+                if tool_stream is not None:
+                    for _frag in tool_stream.flush():
+                        yield "data: " + json.dumps(
+                            _batched_tool_payload(
+                                _frag, rid, request.model, in_toks, len(sub.token_ids), sub
+                            )
+                        ) + "\n\n"
+                yield "data: " + json.dumps(
+                    {
+                        "id": rid,
+                        "object": "chat.completion.chunk",
+                        "created": int(time.time()),
+                        "model": request.model,
+                        "choices": [
+                            {
+                                "index": 0,
+                                "delta": {"role": "assistant", "content": ""},
+                                "finish_reason": "stop" if reason == "stop" else "length",
+                            }
+                        ],
+                        "usage": {
+                            "input_tokens": in_toks,
+                            "output_tokens": len(sub.token_ids),
+                            "total_tokens": in_toks + len(sub.token_ids),
+                            "prompt_tps": prompt_tps,
+                            "generation_tps": generation_tps,
+                            "cached_tokens": sub.reused_tokens,
+                        },
+                    }
+                ) + "\n\n"
+                yield "data: [DONE]\n\n"
+                return
+            elif kind == "err":
+                yield "data: " + json.dumps({"error": item[1]}) + "\n\n"
+                return
+    finally:
+        # If we exit early (disconnect / error), make sure the worker evicts us.
+        sub.cancelled = True
+        # Release the eviction pins taken for this turn's cache + prefix.
+        _unmark_inflight(sub.pinned_cache_ids)
+        sub.pinned_cache_ids = []
+        engine = _BATCH_ENGINE
+        _reclaim_mlx_buffer_cache(
+            "batch-turn-end",
+            force=engine is None or engine.is_idle(),
+            log=True,
+        )
+
+
+async def _await_batched_completion(sub: "_Sub") -> None:
+    """Drain a non-streaming BatchEngine sub (currently cache warming)."""
+    completed = False
+    try:
+        while True:
+            item = await sub.queue.get()
+            kind = item[0]
+            if kind == "done":
+                completed = True
+                return
+            if kind == "err":
+                raise RuntimeError(item[1])
+            # Generated token deltas are intentionally discarded: warming
+            # needs one decode step to finalize the cache, not its text.
+    finally:
+        if not completed:
+            sub.cancelled = True
+        _unmark_inflight(sub.pinned_cache_ids)
+        sub.pinned_cache_ids = []
+        engine = _BATCH_ENGINE
+        _reclaim_mlx_buffer_cache(
+            "batch-warm-end",
+            force=engine is None or engine.is_idle(),
+            log=True,
+        )
+
+
+_BATCH_ENGINE: "Optional[BatchEngine]" = None
+
+
+def _get_batch_engine() -> "BatchEngine":
+    global _BATCH_ENGINE
+    if _BATCH_ENGINE is None:
+        lm = getattr(MODEL, "language_model", MODEL)
+        tok = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+        # mlx_lm models return raw logits — no LanguageModelOutput to unwrap.
+        engine_model = lm if _TEXT_TOWER == "mlx_lm" else _UnwrapLM(lm)
+        _BATCH_ENGINE = BatchEngine(engine_model, tok, ARGS.max_concurrency)
+    return _BATCH_ENGINE
+
+
+def _flush_all_to_disk() -> int:
+    """Write every in-memory entry to disk. Returns the count saved.
+
+    Called from the SIGTERM handler and the FastAPI shutdown event so a
+    graceful supervisor stop preserves every warm session for the next
+    boot. SIGKILL still loses everything (no handler runs); that's the
+    pre-persistence cost as a worst case.
+    """
+    if not _disk_enabled():
+        return 0
+    saved = 0
+    clean = 0
+    started = time.perf_counter()
+    for cache_id, entry in list(_CACHE.items()):
+        if not entry.dirty:
+            clean += 1
+            continue
+        if cache_persist.save_cache(
+            PERSIST_DIR, MODEL_FINGERPRINT, cache_id, entry.state  # type: ignore[arg-type]
+        ):
+            entry.dirty = False
+            saved += 1
+    if saved or clean:
+        print(
+            f"[cache] flushed {saved} entries to disk ({clean} already persisted) "
+            f"in {time.perf_counter() - started:.1f}s",
+            flush=True,
+        )
+    _enforce_disk_budget()
+    return saved
+
+
+# ───────── HTTP request/response models ─────────
+
+
+class ChatMessageReq(BaseModel):
+    role: str
+    content: Any
+    images: Optional[List[str]] = None  # raw base64, never a file path or URL
+    tool_call_id: Optional[str] = None
+    # Undeclared fields are dropped by pydantic, and a dropped
+    # `tool_calls` silently costs the model its own tool-call history:
+    # Gemma's template renders a `role: "tool"` message only when it can
+    # pair it with the preceding assistant call, so without this the
+    # whole call/result exchange vanishes from the prompt. Wild-caught
+    # on gemma4-26b-q4 — 62% of turns re-sent a prompt just ~5 tokens
+    # longer than the last, the model never saw a single tool result,
+    # and it re-issued the same call until the repeat-guard killed the
+    # turn. Qwen masked it: Hermes templates render a bare tool message.
+    tool_calls: Optional[List[Dict[str, Any]]] = None
+
+
+class ChatRequest(BaseModel):
+    model: str
+    messages: List[ChatMessageReq]
+    stream: bool = True
+    tools: Optional[List[Dict[str, Any]]] = None
+    # Gezel extension. When set, the server preserves the post-generation
+    # cache state under this key. When unset, behaves like upstream
+    # mlx_vlm.server (cache cleared after the stream).
+    cache_id: Optional[str] = None
+    # Cross-session prefix sharing (Phase 1.3). When the session-specific
+    # `cache_id` misses, the server falls back to loading this entry —
+    # typically a per-gezel hash of the stable system prompt — and
+    # registers it under the session's cache_id as a starting state.
+    # First chat turn for a new session inherits the warm prefix
+    # instead of cold-prefilling 2 KB of system prompt.
+    prefix_cache_id: Optional[str] = None
+    # Layered prefix sharing (flag `layeredPrefixCache`). An ordered,
+    # most-specific-first list of prefix ids — typically
+    # `[prefix-gp-<hash>, prefix-gezel-<hash>]`. On a session miss the
+    # server tries each in order and seeds from the first hit, so a new
+    # session inherits the warmest matching layer (gezel+project > gezel).
+    # When set, takes precedence over the singular `prefix_cache_id`.
+    prefix_cache_ids: Optional[List[str]] = None
+    # Length in CHARACTERS of the shared band inside messages[0].content —
+    # the leading run sibling sessions of the same (gezel, project) render
+    # identically. When set, the engine plants its prompt-snapshot cut at
+    # that boundary and publishes the SHORT snapshot as the prefix entry
+    # instead of the session's full post-turn state.
+    #
+    # Why this matters here and not on llama.cpp: 48 of qwen3.8's 64 layers
+    # are linear-attention `ArraysCache` holding a fixed-size recurrent
+    # state with no `trim()`, so a prefix entry LONGER than the shared head
+    # is not a partial hit — `seed_from_state` falls to `fresh-untrimmable`
+    # and the whole prompt is re-prefilled. Short entries are the only
+    # reusable shape. See ADR 0010.
+    stable_prefix_chars: Optional[int] = None
+    # Generation knobs (subset of mlx-vlm's). Not all are honored today —
+    # we just need enough to run sensible defaults.
+    max_tokens: Optional[int] = None
+    temperature: Optional[float] = None
+    top_p: Optional[float] = None
+    top_k: Optional[int] = None
+    repetition_penalty: Optional[float] = None
+    repetition_context_size: Optional[int] = None
+    # `tuning.reasoning.thinkingBudget` — enforced here because MLX has no
+    # `--reasoning-budget` equivalent; see think_budget.py.
+    max_thinking_tokens: Optional[int] = None
+    # Decode-time tool-call grammar (gezel extension). When set, the
+    # server builds an llguidance logits processor that constrains the
+    # function name of any tool call to the known tools — see
+    # `build_tool_grammar_processor`. Shape: {"format": "...", "mode": "name-only"}.
+    tool_grammar: Optional[Dict[str, Any]] = None
+    # Per-request circuit breaker for a session that observed the assisted
+    # speculative path violate a required-argument grammar. The TS provider
+    # latches this after the first impossible call and leaves the drafter
+    # resident for unaffected sessions; only this request takes the ordinary
+    # BatchGenerator path. This is the request-scoped equivalent of the
+    # operator's GEZEL_MLX_SPEC=greedy-only A/B arm.
+    disable_speculation: bool = False
+    # Who is waiting on this turn: "interactive" (a person typed it) or
+    # "background" (task steps, nudges, warms). A running wave made only of
+    # background work parks for a waiting interactive request; absent reads
+    # as interactive so an older daemon never gets its turns parked.
+    priority: Optional[str] = None
+    # Per-request chat-template override (gezel extension). Swaps the
+    # model's stored Jinja template for a curated one without reinstalling.
+    chat_template_override: Optional[str] = None
+    # Template variables the caller wants bound when rendering the prompt —
+    # `enable_thinking`, `reasoning_effort`, and friends. llama-server has
+    # honored this field for a while and the MLX provider has always SENT it;
+    # until it was declared here pydantic dropped it silently, so every
+    # constrained turn that asked for thinking to be off still rendered a
+    # prompt ending in `<think>`. That one missing field was the bulk of the
+    # measured MLX/llama-cpp divergence — see `_build_prompt`.
+    chat_template_kwargs: Optional[Dict[str, Any]] = None
+
+
+class CacheStatsEntry(BaseModel):
+    cache_id: str
+    token_count: int
+    est_bytes: int
+    last_used_ts: float
+
+
+# ───────── FastAPI app ─────────
+
+
+app = FastAPI(title="gezel-mlx-server")
+
+
+@app.on_event("shutdown")
+def _on_shutdown() -> None:
+    """uvicorn fires this on SIGTERM (graceful) before exiting.
+
+    Persisting here means a clean supervisor stop preserves every warm
+    session — the user comes back after lunch, the prefix loads from
+    disk in <200ms instead of paying a full re-prefill.
+    """
+    _flush_all_to_disk()
+
+
+# SIGTERM handler as a belt-and-braces complement to FastAPI's
+# `on_event('shutdown')`. uvicorn's signal handling installs its own
+# trap before ours, so the on_event hook is the primary path; this
+# fallback runs only when uvicorn's handler doesn't reach the
+# shutdown event (some kill scenarios race the handler installation).
+def _sigterm_handler(_signum: int, _frame: Any) -> None:
+    _flush_all_to_disk()
+    # Re-raise default behavior — let uvicorn's own shutdown finish.
+    sys.exit(0)
+
+
+try:
+    import signal as _signal
+
+    _signal.signal(_signal.SIGTERM, _sigterm_handler)
+except (ValueError, AttributeError):
+    # `signal.signal` only works on the main thread; if we somehow
+    # aren't there, on_event is still wired and that's the common path.
+    pass
+
+
+@app.get("/health")
+async def health() -> JSONResponse:
+    """Same shape mlx_vlm.server emits — keeps our supervisor probe happy."""
+    return JSONResponse({"status": "ok", "vision": _VISION_ENABLED})
+
+
+@app.post("/admin/flush")
+async def admin_flush(settle_ms: int = 0) -> JSONResponse:
+    """Persist every in-memory cache entry to disk without dropping it.
+
+    Called by the gezel supervisor's Stage-1 idle freeze: the user has
+    been quiet long enough that we should defensively save state, but
+    not so long that we want to evict the model from unified memory.
+    If the user comes back during the freeze window the in-memory
+    cache is still warm; if they DON'T come back and Stage 2 SIGTERMs
+    the process, the freeze flush already wrote everything out so the
+    next boot resumes cleanly.
+
+    Also called right before a deliberate stop, with `settle_ms`: the
+    daemon cancels its in-flight turns first, and a cancelled turn files
+    its prompt snapshot only when the worker next steps. Waiting (bounded)
+    for the engine to drain is what gets that snapshot into this flush
+    instead of losing it to the SIGKILL that follows SIGTERM.
+    """
+    engine = _BATCH_ENGINE
+    settled = True
+    if settle_ms > 0 and engine is not None:
+        deadline = time.time() + min(int(settle_ms), 60_000) / 1000.0
+        while not engine.is_idle() and time.time() < deadline:
+            await asyncio.sleep(0.05)
+        settled = engine.is_idle()
+    saved = _flush_all_to_disk()
+    return JSONResponse({"flushed": saved, "settled": settled})
+
+
+@app.get("/v1/cache/stats")
+async def cache_stats() -> List[CacheStatsEntry]:
+    """In-memory entries only.
+
+    Intentionally excludes disk-only persisted entries — the controller
+    polling this endpoint manages the in-memory budget against the
+    in-engine LRU. Disk persistence is a separate, larger budget
+    enforced internally by _enforce_disk_budget on each save. Surfacing
+    disk entries here would force the controller to evict them as if
+    they were live, defeating the persistence behavior.
+    """
+    return [
+        CacheStatsEntry(
+            cache_id=cid,
+            token_count=len(entry.state.token_ids or []),
+            est_bytes=entry.est_bytes,
+            last_used_ts=entry.last_used_ts,
+        )
+        for cid, entry in _CACHE.items()
+    ]
+
+
+@app.delete("/v1/cache/{cache_id}")
+async def cache_evict(cache_id: str) -> JSONResponse:
+    """Drop an entry from both in-memory cache and disk persistence.
+
+    Disk + memory in lockstep is intentional: a controller-driven evict
+    means the operator (or an LRU policy) decided this session's cache
+    is no longer useful. Leaving the disk entry behind would defeat
+    the eviction (next access would re-load it). Deleting both is the
+    "clean slate" the controller intends.
+    """
+    in_memory = cache_id in _CACHE
+    if in_memory:
+        del _CACHE[cache_id]
+    on_disk = (
+        cache_persist.delete_cache(PERSIST_DIR, MODEL_FINGERPRINT, cache_id)  # type: ignore[arg-type]
+        if _disk_enabled()
+        else False
+    )
+    if in_memory or on_disk:
+        print(
+            f"[cache] evicted cache_id={cache_id} (memory={in_memory} disk={on_disk})",
+            flush=True,
+        )
+    return JSONResponse({"evicted": in_memory or on_disk, "cache_id": cache_id})
+
+
+class CacheWarmRequest(BaseModel):
+    cache_id: str
+    messages: List[ChatMessageReq]
+    model: Optional[str] = None
+    # A warmed prefix is only reusable if it is a genuine token-prefix of the
+    # real turns that follow, which means rendering through the SAME template
+    # branch they do. Without `tools` this path fell to the non-tokenizer
+    # render, and on Qwen — which puts the tool block at the top of the system
+    # message — the warmed entry shared 3 tokens with every real prompt. Every
+    # prefix-hit was a false hit: disk load, eviction pin, then a full
+    # prefill anyway.
+    tools: Optional[List[Dict[str, Any]]] = None
+    chat_template_kwargs: Optional[Dict[str, Any]] = None
+    # When True, write the warmed entry to disk immediately (don't wait
+    # for eviction or shutdown). Used by the gezel-prefix warming flow
+    # so subsequent sessions can load the prefix on miss without
+    # depending on supervisor lifecycle. Default False matches the
+    # session-warm semantics (in-memory only until evicted).
+    persist: bool = False
+
+
+@app.post("/v1/cache/warm")
+async def cache_warm(req: CacheWarmRequest) -> JSONResponse:
+    """Warm a cache through the same snapshot-capable BatchGenerator as chat.
+
+    The older serial warm saved post-generation state. That is reusable for
+    full-attention caches (they can trim back to the shared prefix), but not for
+    wrapped sliding-window caches: their synthetic user/generated tail cannot
+    be rewound, so the real first turn paid a full prefill anyway.
+
+    System-only warms still need a synthetic user for Qwen-family templates.
+    We render a second, deliberately different synthetic user and take their
+    token LCP minus the normal safety margin. That gives BatchEngine an exact
+    snapshot boundary inside the stable system prefix, before either fake user
+    can contaminate an untrimmable entry.
+    """
+    if any(m.images for m in req.messages):
+        raise HTTPException(status_code=422, detail="Vision requests cannot warm text-prefix caches")
+    synthetic_user = not any(m.role == "user" for m in req.messages)
+    warm_messages = list(req.messages)
+    if synthetic_user:
+        warm_messages.append(ChatMessageReq(role="user", content="hi"))
+
+    prompt_text = _build_prompt(
+        warm_messages, req.tools, None, req.chat_template_kwargs
+    )
+    tok = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+    prompt_tokens = [int(t) for t in tok.encode(prompt_text)]
+    snapshot_target = None
+    if synthetic_user:
+        alternate_messages = list(req.messages) + [
+            ChatMessageReq(role="user", content="__gezel_cache_boundary_probe__")
+        ]
+        alternate_text = _build_prompt(
+            alternate_messages, req.tools, None, req.chat_template_kwargs
+        )
+        alternate_tokens = [int(t) for t in tok.encode(alternate_text)]
+        target = cache_seed.stable_snapshot_boundary(
+            prompt_tokens, alternate_tokens, _SNAPSHOT_BOUNDARY_MARGIN
+        )
+        snapshot_target = target
+
+    # A warm entry must be stable on EVERY axis the real turns vary, not just
+    # the synthetic user. Measured: warming stopped at the synthetic-user
+    # boundary (2,509 tokens) while a constrained turn diverges 27 tokens
+    # earlier at the reasoning preamble (2,482) — and an untrimmable cache
+    # cannot rewind 27 tokens any more than it can rewind 2,000, so every
+    # prefix hit was still a full prefill. Take the tighter of the two.
+    reasoning_target = _reasoning_stable_boundary(req, prompt_tokens)
+    if reasoning_target:
+        snapshot_target = (
+            reasoning_target if not snapshot_target else min(snapshot_target, reasoning_target)
+        )
+        if snapshot_target == 0:
+            print(
+                f"[cache] no stable synthetic-user boundary for "
+                f"cache_id={req.cache_id}; using normal finish save",
+                flush=True,
+            )
+
+    warm_request = ChatRequest(
+        model=req.model or str(ARGS.model),
+        messages=warm_messages,
+        stream=True,
+        cache_id=req.cache_id,
+        max_tokens=1,
+        temperature=0.0,
+        # Nobody reads a warm's output; it must never hold a person up.
+        priority=wave_policy.BACKGROUND,
+    )
+    cache_state, _, _ = _resolve_cache_state(warm_request)
+    prior_tokens = len(cache_state.token_ids or [])
+    warm_roster_fp = _roster_fp_of(req.tools)
+
+    # A warm must never race a live turn for the same cache_id. Both write
+    # the same entry, so whichever finishes last wins, and the warm's prompt
+    # is by construction the SHORT one. Observed: a 520-token warm landed on
+    # a session mid-turn and replaced its 121,508-token entry with 521
+    # tokens; the in-flight turn had already seeded so it survived, but the
+    # next turn on that session would have re-prefilled 121k tokens from
+    # scratch. The live turn owns the id — refuse, don't queue.
+    if _INFLIGHT_CACHE_IDS.get(req.cache_id, 0) > 0:
+        print(
+            f"[cache] warm skipped for cache_id={req.cache_id}: a request is already "
+            f"in flight on that cache (prior={prior_tokens})",
+            flush=True,
+        )
+        return JSONResponse(
+            {
+                "warmed": False,
+                "cache_id": req.cache_id,
+                "token_count": prior_tokens,
+                "persisted": False,
+                "skipped": "in-flight",
+            }
+        )
+
+    # The tool roster IS part of the prefix identity — it is rendered into
+    # the prompt ahead of everything else, so warming with a different
+    # roster diverges at token ~1 and the "warm" is a full prefill that
+    # then overwrites a good entry with a useless one. Refuse when the
+    # rosters disagree and there is a real entry to lose. An unknown saved
+    # roster is not a mismatch (nothing recorded yet).
+    saved_roster_fp = _LAST_SAVE_ROSTER.get(req.cache_id)
+    if prior_tokens > 0 and saved_roster_fp and saved_roster_fp != warm_roster_fp:
+        print(
+            f"[cache] warm skipped for cache_id={req.cache_id}: roster mismatch "
+            f"(warm={warm_roster_fp} saved={saved_roster_fp}, prior={prior_tokens}) — "
+            f"warming a different tool shape would churn the whole prefix",
+            flush=True,
+        )
+        return JSONResponse(
+            {
+                "warmed": False,
+                "cache_id": req.cache_id,
+                "token_count": prior_tokens,
+                "persisted": False,
+                "skipped": "roster-mismatch",
+            }
+        )
+
+    sub = _Sub(
+        request=warm_request,
+        prompt_tokens=prompt_tokens,
+        grammar=None,
+        seed_state=cache_state,
+        request_id=f"cachewarm-{uuid.uuid4()}",
+        http_request=None,
+        snapshot_target=snapshot_target,
+        roster_fp=warm_roster_fp,
+    )
+    sub.pinned_cache_ids = _mark_inflight(req.cache_id)
+    sub.prefill_total = len(prompt_tokens)
+    try:
+        _get_batch_engine().submit(sub)
+    except Exception as exc:
+        _unmark_inflight(sub.pinned_cache_ids)
+        sub.pinned_cache_ids = []
+        log_contained_exception("http")
+        raise HTTPException(status_code=500, detail=str(exc))
+    try:
+        await _await_batched_completion(sub)
+    except Exception as exc:
+        log_contained_exception("http")
+        raise HTTPException(status_code=500, detail=str(exc))
+
+    cache_state = sub.saved_cache_state
+    new_tokens = len(cache_state.token_ids or []) if cache_state is not None else 0
+    persisted = False
+    if cache_state is not None and new_tokens > 0:
+        # Prefix warming sets persist=True so the entry survives the next
+        # supervisor stop and brand-new sessions can load it on first
+        # request. Session warming leaves it in-memory only; the next
+        # eviction will write through.
+        if req.persist and _disk_enabled():
+            persisted = cache_persist.save_cache(
+                PERSIST_DIR, MODEL_FINGERPRINT, req.cache_id, cache_state  # type: ignore[arg-type]
+            )
+            resident = _CACHE.get(req.cache_id)
+            if persisted and resident is not None and resident.state is cache_state:
+                resident.dirty = False
+            _enforce_disk_budget()
+        print(
+            f"[cache] warmed cache_id={req.cache_id} tokens={new_tokens} "
+            f"(prior={prior_tokens}, persist={persisted})",
+            flush=True,
+        )
+    else:
+        print(
+            f"[cache] warm produced no tokens for cache_id={req.cache_id}; "
+            f"leaving prior entry intact (prior={prior_tokens})",
+            flush=True,
+        )
+    return JSONResponse(
+        {
+            "warmed": new_tokens > 0,
+            "cache_id": req.cache_id,
+            "token_count": new_tokens,
+            "persisted": persisted,
+        }
+    )
+
+
+def _template_tool_calls(tool_calls: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+    """Shape OpenAI-wire tool calls for Jinja chat templates.
+
+    The wire format carries `function.arguments` as a JSON *string*;
+    HF chat templates near-universally expect a *mapping* and iterate it
+    (`arguments | items`, `arguments.path`). Qwen's template raises
+    `TypeError: Can only get item pairs from a mapping` on a string,
+    which `apply_chat_template`'s caller catches — and its last-resort
+    recovery is to re-render with `tools` dropped entirely, so a string
+    here costs the model its whole tool surface rather than one message.
+    Gemma's template tolerates both, which is why this only bites some
+    families. Parse when we can; pass through untouched when we can't
+    (a non-JSON string is still better than an exception).
+    """
+    shaped: List[Dict[str, Any]] = []
+    for call in tool_calls:
+        if not isinstance(call, dict):
+            shaped.append(call)
+            continue
+        fn = call.get("function")
+        args = fn.get("arguments") if isinstance(fn, dict) else None
+        if not isinstance(args, str):
+            shaped.append(call)
+            continue
+        try:
+            parsed = json.loads(args)
+        except (ValueError, TypeError):
+            shaped.append(call)
+            continue
+        if not isinstance(parsed, dict):
+            shaped.append(call)
+            continue
+        shaped.append({**call, "function": {**fn, "arguments": parsed}})
+    return shaped
+
+
+_TOOL_LINKAGE_PROBE: Dict[str, bool] = {}
+_LINKAGE_PROBE_MARK = "__gezel_tool_linkage_probe__"
+
+
+def _tool_linkage_probe_messages(linked: bool) -> List[Dict[str, Any]]:
+    call = {
+        "id": "probe1",
+        "type": "function",
+        "function": {"name": "probe_tool", "arguments": {"q": "x"}},
+    }
+    assistant: Dict[str, Any] = {"role": "assistant", "content": ""}
+    result: Dict[str, Any] = {"role": "tool", "content": _LINKAGE_PROBE_MARK}
+    if linked:
+        assistant["tool_calls"] = [call]
+        result["tool_call_id"] = "probe1"
+    return [{"role": "user", "content": "go"}, assistant, result]
+
+
+def _template_needs_tool_linkage(tokenizer: Any, chat_template_override: Any) -> bool:
+    """Does this model's template DROP a tool result unless the assistant
+    call it answers is linked to it?
+
+    Gemma renders a `role: "tool"` message only as the response half of a
+    call/response pair, so without the linkage the entire exchange vanishes
+    and the model never sees a single tool result. Hermes-style templates
+    (Qwen et al.) render a bare tool message just fine.
+
+    That difference is why this is probed rather than always-on. Feeding
+    Qwen the linkage is *more* correct on paper but measurably worse in
+    practice: it starts seeing its own past `<tool_call>` blocks echoed in
+    the transcript and imitates them — core-suite MLX went 8/11 -> 5/11,
+    tool calls 457 -> 753, three fast passes turning into slow model-stuck
+    retry loops. So we only add the linkage where its absence actually
+    costs the model information, and leave every other template on the byte-
+    identical legacy shape.
+
+    Probed once per template: render a marker-bearing tool result with and
+    without linkage; if the bare render loses the marker, linkage is
+    required. Failures answer "no" — the legacy shape is the safe default.
+    """
+    key = str(chat_template_override) if chat_template_override else "<model>"
+    cached = _TOOL_LINKAGE_PROBE.get(key)
+    if cached is not None:
+        return cached
+    probe_tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "probe_tool",
+                "description": "probe",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"q": {"type": "string"}},
+                },
+            },
+        }
+    ]
+    kwargs: Dict[str, Any] = dict(
+        tools=probe_tools, add_generation_prompt=True, tokenize=False
+    )
+    if chat_template_override:
+        kwargs["chat_template"] = chat_template_override
+    needs = False
+    try:
+        bare = tokenizer.apply_chat_template(
+            _tool_linkage_probe_messages(False), **kwargs
+        )
+        linked = tokenizer.apply_chat_template(
+            _tool_linkage_probe_messages(True), **kwargs
+        )
+        needs = (_LINKAGE_PROBE_MARK not in bare) and (_LINKAGE_PROBE_MARK in linked)
+    except Exception as exc:
+        print(
+            f"[tool-prompt] tool-linkage probe failed ({exc}); "
+            f"using legacy message shape",
+            flush=True,
+        )
+        needs = False
+    _TOOL_LINKAGE_PROBE[key] = needs
+    print(
+        f"[tool-prompt] tool-linkage probe: template "
+        f"{'REQUIRES' if needs else 'does not need'} assistant/tool pairing"
+        f"{' — passing tool_calls + tool_call_id through' if needs else ''}",
+        flush=True,
+    )
+    return needs
+
+
+# Template variables naming reasoning DEPTH rather than the on/off switch.
+# Mirrors `REASONING_DEPTH_TEMPLATE_KWARGS` in providers/reasoning-depth.ts;
+# kept in sync by reasoning-template-kwargs-wiring.test.ts.
+_REASONING_DEPTH_KWARGS = frozenset({"reasoning_effort", "reasoning_strength"})
+
+
+def _reasoning_stable_boundary(request, prompt_tokens) -> Optional[int]:
+    """Last prompt position that does not depend on this turn's reasoning mode.
+
+    Gezel flips `enable_thinking` between constrained and ordinary turns, and
+    the template renders a reasoning preamble whose presence follows it. With
+    the preamble relocated below the stable content that difference sits at
+    the END of the system block — close enough to the whole prompt to look
+    harmless, and fatal anyway, because a hybrid (linear-attention) cache
+    cannot be trimmed back by even one token.
+
+    Returns an absolute boundary to snapshot at, or None to keep the default
+    end-of-prompt behaviour (correct for models whose cache CAN trim).
+    """
+    if not _needs_reasoning_stable_snapshot():
+        return None
+    kwargs = dict(request.chat_template_kwargs or {})
+    flipped = dict(kwargs)
+    flipped["enable_thinking"] = not bool(kwargs.get("enable_thinking", True))
+    try:
+        alternate = _build_prompt(
+            request.messages, request.tools, request.chat_template_override, flipped
+        )
+        tok = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+        alternate_tokens = [int(t) for t in tok.encode(alternate)]
+    except Exception as exc:  # noqa: BLE001 — never fail a turn over a boundary hint
+        print(f"[cache] reasoning-stable boundary unavailable: {exc}", flush=True)
+        return None
+    boundary = cache_seed.stable_snapshot_boundary(
+        prompt_tokens, alternate_tokens, _SNAPSHOT_BOUNDARY_MARGIN
+    )
+    # 0 disables capture entirely in plan_snapshot_segments. If the two modes
+    # diverge immediately there is nothing stable to save, and the default
+    # end-of-prompt snapshot would just re-save an unusable entry.
+    return boundary or None
+
+
+def _needs_reasoning_stable_snapshot() -> bool:
+    """True when a volatile tail in the saved state costs the ENTIRE entry.
+
+    Deliberately the same predicate the engine already uses to decide whether
+    to capture prompt snapshots at all (`prompt_snapshot=y` in the boot log),
+    rather than a second hand-rolled cache probe: an earlier version called
+    `all_trimmable(MODEL.make_cache())` and defaulted to False on any problem,
+    which silently disabled this fix on precisely the untrimmable models it
+    exists for. One predicate, one answer.
+    """
+    if not _STABLE_PREFIX_ENABLED:
+        return False
+    global _REASONING_SNAPSHOT_NEEDED
+    if _REASONING_SNAPSHOT_NEEDED is None:
+        # Probe the SAME object the engine probes — `_get_batch_engine`
+        # builds `BatchEngine(_UnwrapLM(lm))`, and the VLM wrapper's
+        # `make_cache` is not the language model's. Probing the wrapper
+        # reported a trimmable cache for a model that logs
+        # `prompt_snapshot=y` and `fresh-untrimmable` in the same run.
+        _REASONING_SNAPSHOT_NEEDED = cache_seed.probe_needs_prompt_snapshot(
+            getattr(MODEL, "language_model", MODEL)
+        )
+        print(
+            f"[cache] reasoning-stable snapshot "
+            f"{'ENABLED (cache cannot be rewound)' if _REASONING_SNAPSHOT_NEEDED else 'not needed (trimmable cache)'}",
+            flush=True,
+        )
+    return _REASONING_SNAPSHOT_NEEDED
+
+
+_REASONING_SNAPSHOT_NEEDED = None
+
+
+def _build_prompt(
+    messages: List[ChatMessageReq],
+    tools: Optional[List[Dict[str, Any]]] = None,
+    chat_template_override: Optional[str] = None,
+    chat_template_kwargs: Optional[Dict[str, Any]] = None,
+) -> str:
+    """Apply the model's chat template to OpenAI-shape messages.
+
+    When `tools` is provided, route through the tokenizer's
+    `apply_chat_template(..., tools=...)` directly. Qwen's tokenizer
+    (and most tool-trained models) renders the tools list using the
+    model's *canonical* tool-use template — this is what tells the
+    model "use `<tool_call>{json}</tool_call>` envelopes" reliably.
+    Without this, the tools array gets dropped on the floor and the
+    model improvises a tool-call format from training memory.
+
+    Failure handling peels off optional kwargs one at a time rather than
+    silently abandoning `tools` on the first error: `enable_thinking` is
+    a Qwen3-only kwarg, so a template that rejects it must NOT cost us the
+    tool rendering. Dropping `tools` is the last resort and is logged
+    loudly (it directly drives tool-call mangling, so it should surface in
+    eval mechanism metrics rather than hide behind salvage).
+
+    `chat_template_override`, when set, swaps the model's stored Jinja
+    template for a curated one (per-family template fix) without
+    reinstalling — used for the A/B and as an escape hatch for models
+    with a known-bad embedded template.
+
+    Vision messages keep structured image placeholders in the canonical
+    template; decoded pixels are supplied separately to mlx_vlm.
+    """
+    # Text-only models (e.g. laguna): load_processor returns the bare
+    # tokenizer itself — no `.tokenizer` attribute. Falling back to
+    # PROCESSOR keeps the canonical tools render; without it this
+    # function silently drops `tools` via the text-only path below.
+    _probe_tokenizer = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+    link_tools = (
+        any(m.tool_calls or m.tool_call_id for m in messages)
+        and _probe_tokenizer is not None
+        and hasattr(_probe_tokenizer, "apply_chat_template")
+        and _template_needs_tool_linkage(_probe_tokenizer, chat_template_override)
+    )
+    raw = []
+    for m in messages:
+        entry: Dict[str, Any] = {"role": m.role, "content": vision_inputs.message_content(m)}
+        # Carry the tool-call linkage through to the template, but ONLY for
+        # templates that actually need it — see _template_needs_tool_linkage.
+        # Templates that render a bare tool message stay byte-identical to
+        # the legacy shape (which also keeps their prompt caches valid).
+        if link_tools:
+            if m.tool_calls:
+                entry["tool_calls"] = _template_tool_calls(m.tool_calls)
+            if m.tool_call_id:
+                entry["tool_call_id"] = m.tool_call_id
+        raw.append(entry)
+
+    tokenizer = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+    use_tokenizer = (
+        tokenizer is not None
+        and hasattr(tokenizer, "apply_chat_template")
+        # Connected apps often send no tools or images. Their reasoning
+        # controls still need the canonical renderer; the text-only fallback
+        # drops these kwargs and silently reopens Qwen's thinking block.
+        and (
+            bool(tools)
+            or bool(chat_template_override)
+            or bool(chat_template_kwargs)
+            or any(m.images for m in messages)
+        )
+    )
+    if use_tokenizer:
+        base_kwargs: Dict[str, Any] = dict(
+            tools=tools,
+            add_generation_prompt=True,
+            tokenize=False,
+        )
+        # A caller-supplied override always wins — it is the escape hatch for
+        # a known-bad embedded template and must not be silently second-
+        # guessed. Otherwise use the prefix-stable rewrite when one verified.
+        if chat_template_override:
+            base_kwargs["chat_template"] = chat_template_override
+        elif STABLE_CHAT_TEMPLATE:
+            base_kwargs["chat_template"] = STABLE_CHAT_TEMPLATE
+        # Reasoning defaults ON (Qwen3 reasoning forced inside <think>), then
+        # the caller's template vars override it. The default has to stay True:
+        # every model whose template reads the switch was tuned here, and
+        # flipping the default would silently disable reasoning for ordinary
+        # turns rather than only the constrained ones the caller names.
+        template_vars: Dict[str, Any] = {"enable_thinking": True}
+        if chat_template_kwargs:
+            template_vars.update(chat_template_kwargs)
+        # Attempt order, most-capable first. The DEPTH dials get their own
+        # rung above the switch because they are the ones that reject values:
+        # Qwen 3.8's jinja `raise_exception`s on any effort outside
+        # {xhigh, medium, low}, and a template raises rather than returning —
+        # so a bad depth value must not cost us `enable_thinking`, which is
+        # worth far more (it decides whether the turn reasons at all).
+        #   1. + all template vars   (switch + caller's depth)
+        #   2. - depth dials         (template rejected a depth VALUE)
+        #   3. - template vars       (template reads none of them)
+        #   4. - tools               (template can't render tools at all)
+        #
+        # Each rung catches Exception, not (TypeError, ValueError): a Jinja
+        # TemplateError is neither, and the fallbacks are strictly more
+        # conservative than the rung above, so degrading beats a dead turn.
+        depth_keys = [k for k in template_vars if k in _REASONING_DEPTH_KWARGS]
+        try:
+            return tokenizer.apply_chat_template(raw, **template_vars, **base_kwargs)
+        except Exception as exc_vars:
+            if depth_keys:
+                without_depth = {
+                    k: v for k, v in template_vars.items() if k not in _REASONING_DEPTH_KWARGS
+                }
+                try:
+                    result = tokenizer.apply_chat_template(
+                        raw, **without_depth, **base_kwargs
+                    )
+                    print(
+                        f"[tool-prompt] template rejected reasoning depth "
+                        f"{ {k: template_vars[k] for k in depth_keys} }; retried "
+                        f"with the switch only: {exc_vars}",
+                        flush=True,
+                    )
+                    return result
+                except Exception:
+                    pass
+            try:
+                result = tokenizer.apply_chat_template(raw, **base_kwargs)
+                # Loud, because falling back here silently restores the
+                # template's own reasoning defaults — which is exactly the
+                # divergence this parameter exists to close.
+                print(
+                    f"[tool-prompt] template rejected {sorted(template_vars)}; "
+                    f"retried without them (reasoning falls back to the "
+                    f"template default): {exc_vars}",
+                    flush=True,
+                )
+                return result
+            except Exception as exc_tools:
+                if tools:
+                    print(
+                        f"[tool-prompt] TOOLS DROPPED — apply_chat_template could "
+                        f"not render {len(tools)} tool(s); model will not see the "
+                        f"canonical tool template and may mangle calls: {exc_tools}",
+                        flush=True,
+                    )
+
+    # No tools, override or template controls, no tokenizer template, or tools-render
+    # failed: text-only path. apply_chat_template signature varies across
+    # mlx-vlm versions; the (processor, config, messages) form is stable.
+    # The mlx_lm tower has no vlm processor/config pair — its tokenizer
+    # renders directly (same chat_template.jinja on disk either way).
+    if _TEXT_TOWER == "mlx_lm":
+        tok = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+        return tok.apply_chat_template(raw, tokenize=False, add_generation_prompt=True)
+    config = getattr(MODEL, "config", None) or {}
+    return apply_chat_template(PROCESSOR, config, raw)
+
+
+def _serial_prompt_token_variants(prompt_text: str) -> "List[List[int]]":
+    """Best-effort re-encodings of the prompt for the serial-path guard.
+
+    mlx_vlm tokenizes the prompt itself (processor pipeline, possibly
+    BOS-prefixed), so our own `encode` can differ by exactly a leading
+    BOS. The guard takes the max LCP across both variants, absorbing
+    that edge without replicating the processor. Empty list = cannot
+    encode = the guard fails safe (reset, cold prefill).
+    """
+    tok = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+    try:
+        base = [int(t) for t in tok.encode(prompt_text)]
+    except Exception:  # noqa: BLE001 — no encode, no verification
+        return []
+    variants = [base]
+    bos = getattr(tok, "bos_token_id", None)
+    if isinstance(bos, int) and not isinstance(bos, bool):
+        if base[:1] == [bos]:
+            variants.append(base[1:])
+        else:
+            variants.append([bos] + base)
+    return variants
+
+
+async def _vision_stream(request, http_request, images):
+    """One full-tower vision request. Never share text-prefix KV or speculate."""
+    request_id = f"chatcmpl-{uuid.uuid4()}"
+    generator = None
+    try:
+        prompt = _build_prompt(request.messages, request.tools,
+                               request.chat_template_override, request.chat_template_kwargs)
+        kwargs = {"prefill_step_size": ARGS.prefill_step_size, **_kv_quant_kwargs()}
+        for key in ("max_tokens", "temperature", "top_p", "top_k",
+                    "repetition_penalty", "repetition_context_size"):
+            value = getattr(request, key, None)
+            if value is not None:
+                kwargs[key] = value
+        tokenizer = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+        processors = []
+        if request.tool_grammar:
+            grammar = tool_grammar.build_tool_grammar_processor(
+                tokenizer, request.tools, request.tool_grammar)
+            if grammar is not None:
+                processors.append(grammar)
+        if request.max_thinking_tokens:
+            processor = think_budget.build_think_budget_processor(
+                tokenizer, request.max_thinking_tokens,
+                opens_in_think=prompt.rstrip().endswith("<think>"))
+            if processor is not None:
+                processors.append(processor)
+        if processors:
+            kwargs["logits_processors"] = processors
+        last = None
+        async with _get_generation_lock():
+            if await http_request.is_disconnected():
+                return
+            print(f"[vision] start request={request_id} images={len(images)} cache=disabled", flush=True)
+            with vision_inputs.isolated_position_state(MODEL):
+                generator = stream_generate(model=MODEL, processor=PROCESSOR,
+                                            prompt=prompt, image=images, **kwargs)
+                try:
+                    for chunk in generator:
+                        if await http_request.is_disconnected():
+                            return
+                        if chunk is None or not hasattr(chunk, "text"):
+                            continue
+                        last = chunk
+                        text = _scrub_leaked_markers(chunk.text)
+                        if text:
+                            yield "data: " + json.dumps({
+                                "id": request_id, "object": "chat.completion.chunk",
+                                "model": request.model,
+                                "choices": [{"index": 0, "delta": {"content": text}, "finish_reason": None}],
+                            }) + "\n\n"
+                        await asyncio.sleep(0)
+                finally:
+                    generator.close()
+                    generator = None
+                    _reclaim_mlx_buffer_cache("vision-end", force=True, log=True)
+        usage = {"input_tokens": getattr(last, "prompt_tokens", 0),
+                 "output_tokens": getattr(last, "generation_tokens", 0),
+                 "cached_tokens": 0,
+                 "prompt_tps": getattr(last, "prompt_tps", 0),
+                 "generation_tps": getattr(last, "generation_tps", 0)}
+        finish = "length" if request.max_tokens and usage["output_tokens"] >= request.max_tokens else "stop"
+        yield "data: " + json.dumps({
+            "id": request_id, "object": "chat.completion.chunk", "model": request.model,
+            "choices": [{"index": 0, "delta": {}, "finish_reason": finish}], "usage": usage,
+        }) + "\n\n"
+        yield "data: [DONE]\n\n"
+    except Exception as exc:
+        log_contained_exception("vision")
+        yield "data: " + json.dumps({"error": str(exc)}) + "\n\n"
+    finally:
+        if generator is not None:
+            generator.close()
+        for image in images:
+            image.close()
+
+
+@app.post("/v1/chat/completions")
+@app.post("/chat/completions")
+async def chat_completions(request: ChatRequest, http_request: Request):
+    if request.stream is False:
+        # Non-streaming path is rare from gezel's side; reject explicitly
+        # rather than silently behaving differently.
+        raise HTTPException(
+            status_code=501,
+            detail="gezel-mlx-server only supports streaming completions",
+        )
+
+    if any(message.images for message in request.messages):
+        if not _VISION_ENABLED:
+            raise HTTPException(status_code=422, detail="This MLX engine has no enabled vision tower")
+        try:
+            images = vision_inputs.decode_images(request.messages)
+        except Exception as exc:
+            raise HTTPException(status_code=422, detail=f"Invalid image input: {exc}") from exc
+        return StreamingResponse(_vision_stream(request, http_request, images),
+                                 media_type="text/event-stream",
+                                 headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"})
+
+    prompt_text = _build_prompt(
+        request.messages,
+        request.tools,
+        request.chat_template_override,
+        request.chat_template_kwargs,
+    )
+
+    # Build the tool-call grammar logits processor once per request (the
+    # llguidance tokenizer is cached across requests). Any failure returns
+    # None — the turn then runs unconstrained + TS salvage, so this never
+    # breaks a turn.
+    tool_grammar_processor = None
+    if request.tool_grammar:
+        _grammar_tok = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+        tool_grammar_processor = tool_grammar.build_tool_grammar_processor(
+            _grammar_tok, request.tools, request.tool_grammar
+        )
+    think_budget_processor = None
+    if request.max_thinking_tokens:
+        _tb_tok = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+        think_budget_processor = think_budget.build_think_budget_processor(
+            _tb_tok,
+            request.max_thinking_tokens,
+            # Thinking templates open the assistant turn already inside the
+            # think block (`...<think>\n` generation prompt) — the open tag
+            # never appears in the generated stream, so seed the state from
+            # the prompt tail.
+            opens_in_think=prompt_text.rstrip().endswith("<think>"),
+        )
+
+    # Look up or initialize cache state. Cascade:
+    #   1. In-memory hit on cache_id           — best case, no I/O.
+    #   2. Disk hit on cache_id                — supervisor restart or
+    #                                            evicted entry, mmap load.
+    #   3. Disk hit on prefix_cache_id         — new session inheriting a
+    #                                            warmed gezel prefix.
+    #   4. Fresh PromptCacheState              — full prefill required.
+    cache_state, cache_source, matched_prefix = _resolve_cache_state(request)
+
+    if cache_source == "memory":
+        existing = len(cache_state.token_ids or [])
+        print(
+            f"[cache] hit cache_id={request.cache_id} prior_tokens={existing}",
+            flush=True,
+        )
+    elif cache_source == "disk":
+        existing = len(cache_state.token_ids or [])
+        print(
+            f"[cache] disk-hit cache_id={request.cache_id} prior_tokens={existing}",
+            flush=True,
+        )
+    elif cache_source == "prefix":
+        existing = len(cache_state.token_ids or [])
+        print(
+            f"[cache] prefix-hit cache_id={request.cache_id} "
+            f"prefix={matched_prefix} prior_tokens={existing}",
+            flush=True,
+        )
+    elif request.cache_id:
+        print(f"[cache] miss cache_id={request.cache_id}", flush=True)
+
+    # Pin this turn's session + seed-prefix entries from eviction for the life
+    # of the request, so a concurrent save/evict can't drop the KV state this
+    # turn is about to reuse out from under it (released in each stream
+    # iterator's finally).
+    pinned_cache_ids = _mark_inflight(request.cache_id, matched_prefix)
+    prior_tokens = len(cache_state.token_ids or [])
+
+    request_id = f"chatcmpl-{uuid.uuid4()}"
+
+    # BatchGenerator path: use it even for a singleton queue. Besides
+    # concurrency, this path owns the prompt-boundary snapshot needed to reuse
+    # wrapped/windowed KV caches safely after a tool result makes the next
+    # prompt diverge from the raw generated-token tail. The serial fallback is
+    # retained only for an explicit --max-concurrency 0.
+    if int(getattr(ARGS, "max_concurrency", 1) or 0) >= 1:
+        _engine = _get_batch_engine()
+        _tok = getattr(PROCESSOR, "tokenizer", None) or PROCESSOR
+        prompt_tokens = [int(t) for t in _tok.encode(prompt_text)]
+        sub = _Sub(
+            request=request,
+            prompt_tokens=prompt_tokens,
+            grammar=[
+                p for p in (tool_grammar_processor, think_budget_processor) if p is not None
+            ]
+            or None,
+            seed_state=cache_state,
+            request_id=request_id,
+            http_request=http_request,
+            # Save at the last reasoning-mode-invariant position. Without
+            # this the saved entry carries this turn's reasoning preamble,
+            # and the next turn in the other mode diverges from it by ~27
+            # tokens — which an untrimmable hybrid cache cannot rewind, so
+            # the whole multi-thousand-token entry is discarded.
+            snapshot_target=_reasoning_stable_boundary(request, prompt_tokens),
+        )
+        sub.pinned_cache_ids = pinned_cache_ids
+        # Pessimistic placeholder until admission: the authoritative
+        # prefill count comes from the seed plan in _seed_args (the old
+        # entry-size arithmetic here under-reported by ~40× whenever the
+        # seed was rejected, which suppressed the liveness markers during
+        # the exact full re-prefills they exist to surface).
+        sub.prefill_total = len(prompt_tokens)
+        try:
+            _engine.submit(sub)
+        except Exception:
+            _unmark_inflight(pinned_cache_ids)
+            raise
+        return StreamingResponse(
+            _batched_stream_iter(sub),
+            media_type="text/event-stream",
+            headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"},
+        )
+
+    # Defense in depth: the tower selector requires --max-concurrency >= 1
+    # for the mlx_lm text tower, so this serial branch (mlx_vlm
+    # stream_generate) should be unreachable in lm mode. If a future edit
+    # breaks that invariant, fail loudly instead of feeding an mlx_lm
+    # model into mlx_vlm's generate loop.
+    if _TEXT_TOWER == "mlx_lm":
+        raise HTTPException(
+            status_code=501,
+            detail=(
+                "serial generation path is unavailable with the mlx_lm text "
+                "tower; restart without GEZEL_MLX_TEXT_TOWER or with "
+                "--max-concurrency >= 1"
+            ),
+        )
+
+    # Serial-path guard: mlx_vlm's divergent-prefix branch trims by
+    # slicing `keys[:, :, :lcp]`, which silently corrupts wrapped
+    # sliding-window layers (the rotated cache is SHORTER than the
+    # prefix, gets skipped, and its offset disagrees with the trimmed
+    # full-attention layers from then on). When the state can't be
+    # trimmed coherently and the prompt genuinely diverges, a cold
+    # prefill is the only correct outcome — reset before handing the
+    # state to stream_generate. Pure extensions pass through untouched.
+    if cache_state.cache is not None and (cache_state.token_ids or []):
+        _variants = _serial_prompt_token_variants(prompt_text)
+        if cache_seed.serial_reset_needed(cache_state.token_ids, cache_state.cache, _variants):
+            print(
+                f"[cache] serial divergence on untrimmable cache; cold prefill "
+                f"cache_id={request.cache_id}",
+                flush=True,
+            )
+            cache_state = PromptCacheState()
+
+    # IMPORTANT: this MUST be `async def`, not `def`. Starlette iterates
+    # sync generators via `iterate_in_threadpool` which can hop threads
+    # between yields — but MLX's GPU streams are thread-local, so any
+    # `mx.synchronize` call from a different thread than the one that
+    # created the stream raises `RuntimeError: There is no Stream(gpu, 0)
+    # in current thread.` async generators get iterated directly in the
+    # event-loop thread, which keeps the stream + generation co-located.
+    # (The body remains synchronous beyond the `yield`s — `for chunk in
+    # stream_generate(...)` blocks the event loop, same as upstream
+    # mlx_vlm.server's pattern.)
+    async def stream_iter():
+        # Tracks whether the gezel client TCP-disconnected mid-stream
+        # (turn timeout fired, user cancelled, etc.). When set, we
+        # `break` out of `stream_generate` so MLX stops generating
+        # tokens nobody will read — without this, uvicorn's underlying
+        # ASGI write keeps failing per token and floods the supervisor
+        # log with `socket.send() raised exception.` for the whole
+        # remainder of the would-be response.
+        client_gone = False
+        try:
+            generation_kwargs: Dict[str, Any] = {
+                "prompt_cache_state": cache_state,
+                "prefill_step_size": ARGS.prefill_step_size,
+                **_kv_quant_kwargs(),
+            }
+            if request.max_tokens is not None:
+                generation_kwargs["max_tokens"] = request.max_tokens
+            if request.temperature is not None:
+                generation_kwargs["temperature"] = request.temperature
+            if request.top_p is not None:
+                generation_kwargs["top_p"] = request.top_p
+            if request.top_k is not None:
+                generation_kwargs["top_k"] = request.top_k
+            if request.repetition_penalty is not None:
+                generation_kwargs["repetition_penalty"] = request.repetition_penalty
+            if request.repetition_context_size is not None:
+                generation_kwargs["repetition_context_size"] = (
+                    request.repetition_context_size
+                )
+            _lps = [
+                p for p in (tool_grammar_processor, think_budget_processor) if p is not None
+            ]
+            if _lps:
+                # Forwarded through stream_generate(**kwargs) → generate_step,
+                # which accepts `logits_processors`. Grammar constrains the
+                # tool-call function name; think-budget forces `</think>` at
+                # the configured reasoning budget (llama parity).
+                generation_kwargs["logits_processors"] = _lps
+
+            output_text = ""
+
+            tool_stream = _make_tool_stream(request_id, getattr(request, 'tools', None))
+            last_chunk = None
+            # Serialize generation across concurrent requests — see
+            # `_GENERATION_LOCK` doc. Held only across the
+            # `stream_generate` loop, NOT across cache lookup/save, so
+            # request N+1 can prep its prefill state in parallel with
+            # request N's streaming. The lock is fair (asyncio.Lock is
+            # FIFO since Python 3.10) so a long generation can't starve
+            # a queued short one indefinitely.
+            async with _get_generation_lock():
+                for chunk in stream_generate(
+                    model=MODEL,
+                    processor=PROCESSOR,
+                    prompt=prompt_text,
+                    **generation_kwargs,
+                ):
+                    if chunk is None or not hasattr(chunk, "text"):
+                        continue
+                    # Cooperative disconnect check. `is_disconnected()` is
+                    # cheap (just peeks at the ASGI receive queue for an
+                    # `http.disconnect` message) so doing it per token has
+                    # no measurable overhead next to MLX's GPU work.
+                    if await http_request.is_disconnected():
+                        print(
+                            f"[stream] client disconnected mid-generation "
+                            f"after {getattr(chunk, 'generation_tokens', 0)} tokens; "
+                            f"aborting",
+                            flush=True,
+                        )
+                        client_gone = True
+                        break
+                    last_chunk = chunk
+                    # Scrub leaked framing tokens per chunk. Special tokens
+                    # finalize as their own detokenizer segment, so a marker
+                    # never straddles two chunks — a per-chunk scrub is safe.
+                    chunk_text = _scrub_leaked_markers(chunk.text)
+                    if not chunk_text:
+                        continue
+                    output_text += chunk_text
+
+                    # With --stream-tool-calls the raw text is split into
+                    # content and tool-call argument deltas, so the client
+                    # sees a real call boundary instead of markup embedded in
+                    # prose. Without it this is a no-op passthrough and the
+                    # gezel-side salvage path behaves exactly as before.
+                    for _frag in _tool_stream_fragments(tool_stream, chunk_text):
+                        yield "data: " + json.dumps(
+                            _tool_stream_payload(_frag, request_id, request.model, chunk)
+                        ) + "\n\n"
+                    if tool_stream is not None:
+                        continue
+
+                    payload = {
+                        "id": request_id,
+                        "object": "chat.completion.chunk",
+                        "created": int(time.time()),
+                        "model": request.model,
+                        "choices": [
+                            {
+                                "index": 0,
+                                "delta": {"role": "assistant", "content": chunk_text},
+                                "finish_reason": None,
+                            }
+                        ],
+                        "usage": {
+                            "input_tokens": getattr(chunk, "prompt_tokens", 0),
+                            "output_tokens": getattr(chunk, "generation_tokens", 0),
+                            "total_tokens": (
+                                getattr(chunk, "prompt_tokens", 0)
+                                + getattr(chunk, "generation_tokens", 0)
+                            ),
+                            "prompt_tps": getattr(chunk, "prompt_tps", 0.0),
+                            "generation_tps": getattr(chunk, "generation_tps", 0.0),
+                            # Actual prefix reused by mlx-vlm after token
+                            # comparison, not merely the size of the cache
+                            # entry we found by id. This distinguishes a real
+                            # hot continuation from a token-diverged cache
+                            # that still has to re-prefill most of the prompt.
+                            "cached_tokens": getattr(chunk, "cached_tokens", 0),
+                        },
+                    }
+                    yield f"data: {json.dumps(payload)}\n\n"
+
+            if tool_stream is not None and last_chunk is not None:
+
+                for _frag in tool_stream.flush():
+
+                    yield "data: " + json.dumps(
+
+                        _tool_stream_payload(_frag, request_id, request.model, last_chunk)
+
+                    ) + "\n\n"
+
+            if client_gone:
+                # Skip terminal + [DONE] frames — there's nobody to
+                # receive them, and yielding through a closed transport
+                # is what we're trying to avoid. Cache save in `finally`
+                # still runs so the prefill we paid for isn't wasted.
+                return
+
+            # Terminal chunk with finish_reason. Mirrors the upstream
+            # server's shape; gezel's MlxSession reads finish_reason
+            # off the last chunk to decide truncation handling.
+            terminal = {
+                "id": request_id,
+                "object": "chat.completion.chunk",
+                "created": int(time.time()),
+                "model": request.model,
+                "choices": [
+                    {
+                        "index": 0,
+                        "delta": {"role": "assistant", "content": ""},
+                        "finish_reason": "stop",
+                    }
+                ],
+            }
+            if last_chunk is not None:
+                prompt_tokens = getattr(last_chunk, "prompt_tokens", 0)
+                cached_tokens = getattr(last_chunk, "cached_tokens", 0)
+                terminal["usage"] = {
+                    "input_tokens": prompt_tokens,
+                    "output_tokens": getattr(last_chunk, "generation_tokens", 0),
+                    "total_tokens": (
+                        prompt_tokens
+                        + getattr(last_chunk, "generation_tokens", 0)
+                    ),
+                    "prompt_tps": getattr(last_chunk, "prompt_tps", 0.0),
+                    "generation_tps": getattr(last_chunk, "generation_tps", 0.0),
+                    "cached_tokens": cached_tokens,
+                }
+                if request.cache_id:
+                    print(
+                        f"[cache] reuse cache_id={request.cache_id} "
+                        f"cached_tokens={cached_tokens} "
+                        f"prefilled_tokens={max(0, prompt_tokens - cached_tokens)}",
+                        flush=True,
+                    )
+            yield f"data: {json.dumps(terminal)}\n\n"
+            yield "data: [DONE]\n\n"
+
+        except Exception as e:
+            print(f"Error during stream generation: {e}", flush=True)
+            log_contained_exception("stream")
+            yield f"data: {json.dumps({'error': str(e)})}\n\n"
+
+        finally:
+            # The cache `finally` decision:
+            #   - cache_id set + state populated → save (the happy path)
+            #   - cache_id set + state empty (e.g. mid-stream crash before
+            #     stream_generate's update() ran) → DON'T save. Saving an
+            #     empty entry would overwrite any prior good cache and
+            #     guarantee the next turn misses + re-prefills full
+            #   - no cache_id → upstream behavior (drop + clear memory)
+            if request.cache_id:
+                tokens = len(cache_state.token_ids or [])
+                if tokens > 0:
+                    _save_cache(request.cache_id, cache_state)
+                    print(
+                        f"[cache] saved cache_id={request.cache_id} tokens={tokens}",
+                        flush=True,
+                    )
+                    # Seed the shared gezel+project prefix from this real
+                    # session save so sibling sessions inherit [system+tools].
+                    _seed_prefix_from_session(request, cache_state)
+                else:
+                    print(
+                        f"[cache] generation produced no tokens for "
+                        f"cache_id={request.cache_id}; leaving prior entry intact",
+                        flush=True,
+                    )
+            else:
+                print("Stream finished without a prompt-cache save.", flush=True)
+            # Release the eviction pins taken for this turn — AFTER the save
+            # above, so this turn's own _save_cache → _enforce_budget can't
+            # evict the entry it just wrote.
+            _unmark_inflight(pinned_cache_ids)
+            # Agentic tool loops issue multiple HTTP sub-turns inside one user
+            # turn. Reclaim scratch buffers at every one of those boundaries,
+            # including the cache-preserving path above; prompt KV remains an
+            # active allocation and is not touched by clear_cache().
+            _reclaim_mlx_buffer_cache("turn-end", force=True, log=True)
+
+    return StreamingResponse(
+        stream_iter(),
+        media_type="text/event-stream",
+        headers={
+            "Cache-Control": "no-cache",
+            # NOTE: mlx_vlm.server sets `Connection: keep-alive` here,
+            # which our gezel daemon would have to strip when running
+            # over HTTP/2. We're a separate process not on HTTP/2 so
+            # the header is fine to send, but explicitly omit to keep
+            # behavior predictable across transports.
+            "X-Accel-Buffering": "no",
+        },
+    )
+
+
+# ───────── Entry point ─────────
+
+if __name__ == "__main__":
+    import logging
+
+    import uvicorn
+
+    class _DropPollingAccessLogs(logging.Filter):
+        # Supervisor polls /health and the gezel cache adapter polls
+        # /v1/cache/stats every few seconds — at INFO that floods the
+        # parent's [mlx]-prefixed logs. Drop only the 200s for those two
+        # endpoints; anything else (4xx/5xx, other routes) still logs.
+        _NOISY = ("/health", "/v1/cache/stats")
+
+        def filter(self, record: logging.LogRecord) -> bool:
+            args = record.args
+            if not isinstance(args, tuple) or len(args) < 5:
+                return True
+            method, path, _http, status = args[1], args[2], args[3], args[4]
+            return not (
+                method == "GET" and status == 200 and path in self._NOISY
+            )
+
+    logging.getLogger("uvicorn.access").addFilter(_DropPollingAccessLogs())
+
+    uvicorn.run(app, host=ARGS.host, port=ARGS.port, log_level="info")