mlx-bun 0.0.11 → 0.0.12

package/docs/investigations/colibri-oracle-pin.md

@@ -0,0 +1,1284 @@
+# Colibri oracle pin for the native GLM-5.2 port
+
+- Date recorded: 2026-07-21
+- mlx-bun branch: `codex/colibri-glm52-port`
+- Gate: Phase 21 G0, oracle inventory and model-free capture record
+
+This note pins the external behavioral and performance oracle used by
+[`colibri-glm52-port.md`](../design/colibri-glm52-port.md). It does not vendor,
+embed, launch, or ship Colibri. The product remains a native Bun + MLX port.
+
+## Pinned checkout
+
+- Repository: `JustVugg/colibri`
+- Local read-only checkout: `/Users/joshrossi/Code/colibri`
+- Commit: `44e489b196c9b7876b3d37a0570ebf1c6f90f54c`
+- Commit date: `2026-07-21T19:04:16+02:00`
+- Commit subject: `Merge #479: CUDA env integration tests (COLI_CUDA modes + expert budget auto-size)`
+- Branch at inspection: `main`
+- Worktree at inspection: clean, including untracked files
+- Ignored build stamp at inspection: `c/.build-config` absent
+
+The checkout is an executable oracle for formats, token/logit behavior,
+routing, cache traces, MTP accept/reject behavior, and same-machine
+performance. It is not a source dependency. Any result must carry this exact
+commit; `main`, `latest`, or an unrecorded later commit is not equivalent.
+
+## License and attribution policy
+
+The pinned repository contains the Apache License 2.0 in `LICENSE`. It contains
+no tracked `NOTICE` or `COPYRIGHT` file.
+
+The native port should reimplement the documented behavior and cite Colibri in
+the design, user documentation, and third-party attribution. If code or other
+copyrightable material is copied or closely derived, the distributed source
+must also:
+
+1. include a copy of Apache-2.0;
+2. retain applicable copyright, patent, trademark, and attribution notices;
+3. mark modified copied files prominently; and
+4. reproduce any future upstream `NOTICE` content that applies to the copied
+   material.
+
+The Apache license does not grant trademark rights beyond customary
+description of origin. Product naming and artwork are not part of the port.
+
+## Public artifact and disk envelope
+
+The pinned project recommends the public Hugging Face artifact
+[`mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp`](https://huggingface.co/mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp).
+It is the same-artifact oracle target for G0 and the initial native loader.
+
+- Expected converted size: approximately 372 GB.
+- G0 download-time floor: at least 400 GB actually free on fast local storage.
+- Source conversion path: approximately 756 GB of network traffic, processed
+  one roughly 5 GB source shard at a time; the growing output plus one source
+  shard is retained, with a default 20 GB free-space stop margin.
+- Previously measured disk preflight, not reverified by this item: about
+  556 GB unallocated plus 123 GB purgeable, or about 679 GB practically
+  available on the internal SSD on 2026-07-21.
+- The downloaded artifact is pinned locally at revision
+  `3cc8db99b1b13fc79325d987ba3c1c430766b3b8`. The complete metadata manifest
+  is recorded below. The initial audit deferred main-shard hashing to preserve
+  cold state; the G0 closure section records the later complete verification.
+
+The public mirror without the corrected MTP head is not an equivalent
+artifact. The pinned README identifies the correct int8 MTP shard byte sizes
+as:
+
+| Shard | Expected bytes |
+|---|---:|
+| first `out-mtp-*` shard | 3,527,131,672 |
+| second `out-mtp-*` shard | 5,366,238,584 |
+| third `out-mtp-*` shard | 1,065,950,496 |
+
+These sizes are a preflight signal, not a sufficient integrity check. The
+native loader must validate the complete tensor set, names, dtypes, dimensions,
+and byte counts before allocation.
+
+### Post-download artifact audit (2026-07-22)
+
+The artifact downloaded through `mlx-bun get` resolves to:
+
+`/Users/joshrossi/.cache/huggingface/hub/models--mateogrgic--GLM-5.2-colibri-int4-with-int8-mtp/snapshots/3cc8db99b1b13fc79325d987ba3c1c430766b3b8`
+
+- Local `refs/main` and the exact-revision HF API both report
+  `3cc8db99b1b13fc79325d987ba3c1c430766b3b8`.
+- Remote and local inventories match exactly: 150 files and
+  `383,760,077,466` bytes (`357.404 GiB`), with no missing/extra files, broken
+  links, `.incomplete` files, or locks. The canonical
+  `name<TAB>size<TAB>blobId` manifest SHA-256 is
+  `d0a400b49ca018bdaf3b7d98686e5bd9ebad121f79580911fec4fe22196747e2`.
+- The model has 141 consecutive main shards (`out-00000` through
+  `out-00140`) and three MTP shards. Header-only validation parsed all 144
+  safetensors, found no offset/size/JSON errors, and accounted for 118,478
+  tensors. The layer-78 MTP family contains 777 U8 payload tensors, 786 F32
+  scale/dense tensors, all 256 routed experts, and all required attention,
+  shared-expert, `eh_proj`, norm, and shared-head tensors.
+- Direct SHA-256 verification passed for all corrected MTP shards:
+  `dc020ddbb87347f7e6711c9e8cd2715ac79a2a9f2b4599ff11b7980a35e3cf88`,
+  `172b49be499a1070505cd13718c47c82165c663d6422e537b1335aae28c331bf`,
+  and `534a1a2a05188dc372f1e0e4f6d72503cbe86fc9a94f085dc6a6a0941b45975d`.
+  `tokenizer.json` also matches its LFS SHA-256, and all five non-LFS metadata
+  files recompute to their advertised Git blob identities. This was the
+  pre-baseline audit; all 145 LFS payloads were subsequently verified for G0.
+- **DSA is absent from this artifact.** There are zero `out-idx-*` files and
+  zero tensor names containing `indexer`, locally and in the remote revision,
+  even though `config.json` declares the DSA/IndexShare geometry. Pinned
+  Colibri checks for every full-layer `indexer.wq_b.weight` tensor and therefore
+  sets `has_dsa=0`. A baseline on this artifact cannot be called a DSA baseline.
+
+The audit exposed a downloader integrity bug. The current HF `?blobs=true`
+schema returns LFS digests as `lfs.sha256`; mlx-bun modeled the property as
+`lfs.oid`. Consequently, the original 145 LFS transfers used Git `blobId`
+filenames but did not compare their payloads to the advertised SHA-256. The
+local fix normalizes `sha256` (with legacy `oid` compatibility), rejects an
+invalid digest, uses the digest as the content-addressed blob name, and restores
+streaming SHA-256 verification. The regression suite exercises the real field
+name and the corrupt-payload refusal path. At this point the 141 main payloads
+remained unverified because a deliberate full read would populate the
+filesystem cache. They were hashed after the controlled runtime work; the
+closure evidence below records the result.
+
+## Snapshot metadata contract
+
+Runtime-required and behaviorally material files are:
+
+| File | Contract |
+|---|---|
+| `config.json` | Required; supplies model geometry, routing, MLA, DSA, and stop metadata. |
+| `tokenizer.json` | Required for chat/serve; byte-level BPE vocabulary, merges, added tokens, and `special` flags are material. |
+| `tokenizer_config.json` | Preserved by the converter; chat-template/tokenizer policy input for the native port. |
+| `generation_config.json` | Optional to the C loader but authoritative for additional EOS IDs; its EOS set is unioned with `config.json`. |
+| `*.safetensors` | Runtime indexes every safetensors file in the snapshot, independent of filename prefix, up to 512 shards. |
+| `model.safetensors.index.json` | Used by selective conversion/download tooling; not required by the C runtime's directory scan. |
+
+The native GLM config must cover at least:
+
+- `hidden_size`, `num_hidden_layers`, `num_attention_heads`,
+  `n_routed_experts`, `num_experts_per_tok`, `n_shared_experts`;
+- `intermediate_size`, `moe_intermediate_size`,
+  `first_k_dense_replace`;
+- `q_lora_rank`, `kv_lora_rank`, `qk_nope_head_dim`,
+  `qk_rope_head_dim`, `v_head_dim`;
+- `n_group`, `topk_group`, `norm_topk_prob`,
+  `routed_scaling_factor`;
+- `rms_norm_eps`, RoPE parameters, `vocab_size`, and every EOS ID; and
+- `index_topk`, `index_n_heads`, `index_head_dim`, plus either explicit
+  `indexer_types` or the frequency/offset schedule.
+
+GLM-5.2 requires `n_group=1` in the pinned engine. DSA is active only when the
+indexer geometry is valid and every full-indexer layer has its required
+weights. Missing DSA weights disable DSA; the native loader should report this
+as a capability decision rather than silently implying full artifact parity.
+
+## Tensor and shard layout contract
+
+The converter emits three logical shard families:
+
+- `out-*`: main model, including resident tensors and main routed experts;
+- `out-mtp-*`: MTP layer at index `num_hidden_layers`; and
+- `out-idx-*`: DSA lightning-indexer tensors.
+
+The runtime itself keys tensors by name and scans all safetensors files, so the
+prefixes are provenance and tooling conventions rather than routing semantics.
+
+Quantized 2-D tensors are stored as:
+
+- `<name>`: `U8` packed int8, int4, or int2 bytes; and
+- `<name>.qs`: `F32` scales.
+
+Per-row and grouped int4 use the same packed low/high-nibble weight bytes.
+Grouped format is inferred from the scale-array byte count; group sizes in the
+pinned detector are 16, 32, 48, 64, 96, 128, 192, or 256. Norms, the router,
+one-dimensional biases, and `e_score_correction_bias` remain F32. The native
+loader must derive and validate logical `[output,input]` dimensions from the
+config and tensor role instead of trusting a flattened U8 payload.
+
+For each routed expert, `gate_proj`, `up_proj`, and `down_proj` plus their
+scales form one expert load. The converted layout is optimized so the three
+weight matrices can be coalesced into one positioned read and exposed as views
+inside an aligned slab; scales occupy a companion F32 slab. A loader may accept
+a non-coalescible fixture for correctness, but the full artifact path must
+preserve and verify the coalesced layout before claiming I/O parity.
+
+The resident model includes embeddings, LM head, norms, MLA projections, the
+first dense MLP layers, shared experts, and routers. Main routed experts remain
+disk-backed and enter bounded resident slots on demand.
+
+## Tokenizer and stop contract
+
+The pinned tokenizer is byte-level BPE with `ignore_merges=true`, no byte
+fallback, a cl100k-style regex split, and ByteLevel processing without an added
+prefix space. `tokenizer.json` must provide `model.vocab` and `model.merges`.
+
+All added tokens are atomic for encode/decode, but `special:true` has a separate
+serving meaning: control tokens such as user/assistant/observation boundaries
+must stop output, while real content markers such as thinking and tool-call
+tokens may be `special:false` and must remain renderable. The native port must
+union EOS IDs from config and generation config and apply the tokenizer's
+control-token policy without swallowing content tokens.
+
+## MTP artifact contract
+
+MTP is a separate sparse layer at `model.layers.<num_hidden_layers>`. It is
+usable only as a complete set. The pinned runtime's sentinel check covers
+`eh_proj`, `enorm`, `hnorm`, `shared_head.norm`, attention, norms, router,
+shared-expert tensors, and routed-expert endpoints; model initialization then
+loads the complete MTP attention/shared/routed-expert family.
+
+The main conversion skips MTP. A separate `--mtp` pass emits `out-mtp-*` and
+defaults its resident and expert tensors to int8. Per-row int4 is not an
+acceptable substitute: upstream measured roughly 0-4% MTP acceptance because
+`eh_proj`'s asymmetric column scales zero its embedding half. The pinned source
+reports roughly 39-59% acceptance at int8. Group-scaled int4 is a separate
+research option, not the public oracle contract.
+
+The native loader must therefore report MTP as active only after verifying the
+complete int8 MTP set. MTP absence, partial shards, int4 dense MTP tensors, or a
+scheduler mode that disables drafting must be visible in capability and runtime
+telemetry.
+
+## Safe model-free target inventory
+
+No target in this section was executed while creating this note.
+
+| Target | What it covers | Oracle-tree effect |
+|---|---|---|
+| `make -C c check` | Clean, portable CPU build, dependency-free C tests, Python stdlib tests. This is the upstream macOS CI gate. | Deletes/recreates ignored build outputs and writes a build stamp. |
+| `make -C c metal-test` | Int8/int4/int2/F32 matmul, routed SwiGLU, large-batch GEMM, compressed MLA decode, cache writes, and an `S=4` MTP-shaped attention case. | Builds and executes `c/backend_metal_test`. |
+| `python3 tools/convert_fp8_to_int4.py --selftest ...` | Model-free FP8 block-dequant check. | Intended to be output-free, but the pinned CLI needs a dummy `--repo` because it formats `None` before reaching the self-test return. |
+| `python3 tools/convert_fp8_to_int4.py --selftest-nvfp4 ...` | Model-free NVFP4 decode/requant check. | Output-free when dependencies are present; same dummy-`--repo` issue. |
+
+The checkout is a read-only oracle, but both Make targets mutate ignored files
+and the Makefile creates `c/.build-config` during parsing. Do not run them in
+place. G0 item 2 must extract the pinned commit to a temporary directory with
+`git archive`, run both targets there, and capture output verbatim in this note.
+That verifies the exact source without modifying the oracle checkout or its Git
+metadata.
+
+At inventory time, Apple clang 21, arm64 Xcode clang, GNU make, and Homebrew
+libomp were present. System Python lacked Transformers, Torch, safetensors,
+NumPy, and ml-dtypes. The mlx-lm oracle venv had Transformers 5.12,
+safetensors, and NumPy, but lacked Torch and ml-dtypes.
+
+## Tiny-fixture feasibility
+
+Available from the pinned source without full GLM-5.2 weights:
+
+- `make_glm_oracle.py`: seed-1234 tiny GLM-MoE with real MLA/DSA geometry,
+  optional checkpoint-compatible FP8 emission, unfused per-expert tensors, and
+  a 32-position teacher-forcing reference;
+- `make_glm_bench_model.py`: deterministic medium model with configurable
+  output directory and optional FP8 layout;
+- local main and DSA conversion passes over those generated safetensors;
+- deterministic synthetic expert layouts in the io_uring test;
+- deterministic real-shape Metal matmul, routed-SwiGLU, and MLA fixtures; and
+- `ROUTE_TRACE`, which emits per-position/per-layer expert IDs and gate weights
+  without changing model computation.
+
+Limitations that must stay explicit:
+
+1. `make_glm_oracle.py` writes `glm_tiny/` and `ref_glm.json` into its current
+   working directory. Run it from temporary storage, never the oracle checkout.
+2. The generator requires Transformers >=5.11, Torch, and safetensors. The
+   inspected Python environments do not currently satisfy that set; provisioning
+   dependencies is a separate user-authorized decision. No package installation
+   was performed for this inventory.
+3. The generator does not produce `tokenizer.json`; token-ID teacher forcing is
+   feasible, but chat-template parity needs a separately provenance-pinned
+   tokenizer fixture.
+4. The pinned project has no model-free generator for a complete native MTP
+   layer. Its tiny generator covers GLM/MLA/DSA, not the layer-`n_layers` MTP
+   tensor family. G0 cannot honestly claim the requested MTP fixture by merely
+   exporting existing tooling. Add a deterministic derived fixture generator in
+   mlx-bun from this pinned tensor contract, or explicitly extend a temporary
+   copy of the generator; record that derivation in `fixtures/README.md`.
+5. The synthetic expert/io_uring fixture is Linux-only. On macOS, use its
+   recorded byte layout as provenance and validate the port through the Metal
+   fixture and mlx-bun's eventual synthetic expert-file generator.
+
+## Manual G0 baseline checklist
+
+This user-directed machine-local run is complete. CI must never repeat the
+full-model work or download the artifact. The 400 GB figure was a pre-download
+capacity floor, not a requirement to retain 400 GB after the 357 GiB artifact
+itself is present.
+
+- [x] Record the stable post-download APFS capacity snapshot, including hard
+      free space, snapshots, date, and commands.
+- [x] Download the public artifact and record its repository revision, complete
+      file manifest, link identities, and byte sizes.
+- [x] Complete cryptographic payload verification for all 141 main shards.
+- [x] Verify `config.json`, tokenizer metadata, generation metadata, main-shard
+      headers, and the three corrected int8 MTP shards.
+- [x] Supply or explicitly waive complete DSA `out-idx-*` weights. The
+      recommended artifact contains none and therefore runs DSA-disabled.
+- [x] Record machine model, chip, 32 GB memory, macOS/build versions, storage
+      device/filesystem, Colibri commit, Metal build flags, and binary hash.
+- [x] Record the final runtime settings and validate the effective plan for
+      each baseline cell. Preserve the upstream macOS doctor/profiler
+      limitations rather than treating their zero-RAM probe as real hardware.
+- [x] Establish a cleared/quiet machine before the sequence and record
+      pre/post memory pressure, compression, and swap for every cell. A
+      per-cell process-list snapshot was not captured and is explicitly not
+      claimed; unique engine PIDs and the bounded EOF-exit contract are pinned.
+- [x] Run quality-preserving true top-8 defaults. Do not use expert top-p,
+      reduced top-k, `CACHE_ROUTE`, or `EXPERT_BUDGET` for the baseline.
+- [x] Measure cold and warm runs with MTP disabled.
+- [x] Measure cold and warm runs with MTP enabled; record draft proposals,
+      accepted/rejected tokens, acceptance length, tokens/forward, and whether
+      the fixed draft/verify kernel-family contract is active.
+- [x] For every cell capture physical footprint, LRU/pin capacities,
+      pin/LRU/load-miss split, resident expert bytes, disk bytes and disk
+      service/wait time, TTFT, decode tokens/s, compression, and swap. Dense
+      bytes are a fixed verified container/plan constant (`10,877,286,144`),
+      not a per-cell runtime counter; the identical artifact/binary/settings
+      contract is enforced instead of claiming six absent dense-byte samples.
+- [x] Preserve prompts, seeds, generated token IDs/text, raw logs, and command
+      lines. Report medians and run-to-run spread; do not quote loaded-machine
+      or one-off numbers.
+- [x] Compare MTP-on and MTP-off wall time, not only tokens/forward.
+
+This checklist is the G0 exit evidence. Completing this inventory, the
+model-free targets, disk preflight, or fixtures alone does not satisfy G0.
+
+## Provenance commands
+
+Read-only commands used for the pin inspection:
+
+```bash
+git -C /Users/joshrossi/Code/colibri rev-parse HEAD
+git -C /Users/joshrossi/Code/colibri branch --show-current
+git -C /Users/joshrossi/Code/colibri status --porcelain=v1 --untracked-files=all
+git -C /Users/joshrossi/Code/colibri log -1 --format='%H%n%ad%n%s' --date=iso-strict
+git -C /Users/joshrossi/Code/colibri ls-files '*NOTICE*' '*COPYRIGHT*'
+git -C /Users/joshrossi/Code/colibri diff --exit-code -- .
+```
+
+Historical isolated execution pattern used for the now-complete model-free
+suites:
+
+```bash
+COLIBRI_PIN=44e489b196c9b7876b3d37a0570ebf1c6f90f54c
+COLIBRI_TMP="$(mktemp -d)"
+git -C /Users/joshrossi/Code/colibri archive "$COLIBRI_PIN" | tar -x -C "$COLIBRI_TMP"
+make -C "$COLIBRI_TMP/c" check
+make -C "$COLIBRI_TMP/c" metal-test
+```
+
+The complete output was captured before removing the temporary directory, and
+no generated binary was copied into either source checkout. The next section
+records G0 item 2's pass/fail output and toolchain versions.
+
+
+## Isolated model-free suite execution (2026-07-22)
+
+G0 item 2 ran only against an archive of the pinned oracle. Before extraction,
+the oracle checkout was exactly at
+`44e489b196c9b7876b3d37a0570ebf1c6f90f54c`, with empty porcelain output
+and both worktree and index diff checks returning 0. A user-owned
+`mlx-bun get mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp` download was
+active; no user training, inference, benchmark, Colibri test, or other
+Bun/Python MLX compute process was active. This workflow did not start,
+interrupt, or modify the download.
+
+The source was extracted with `git archive` into the validated temporary
+directory
+`/var/folders/t3/bdj253f52sv60p_9rmzm86tr0000gn/T/colibri-g0-pin.XXXXXX.phRD7CaPU1`. It contained no `.git` metadata. Extracted source hashes matched
+the pinned objects:
+
+```text
+c/Makefile  b14f9deaf67adc6285facff41914b4bc64089316c83711ae7c7e7dbb731c814a
+c/glm.c     3be1b4dd663667c8fa2cfbbacdada3e545ff5f924737a0f5d058708d7bc5ad9d
+```
+
+Exact extraction and execution commands:
+
+```bash
+COLIBRI_PIN=44e489b196c9b7876b3d37a0570ebf1c6f90f54c
+COLIBRI_TMP='/var/folders/t3/bdj253f52sv60p_9rmzm86tr0000gn/T/colibri-g0-pin.XXXXXX.phRD7CaPU1'
+git -C /Users/joshrossi/Code/colibri archive "$COLIBRI_PIN" | tar -x -C "$COLIBRI_TMP"
+make -C "$COLIBRI_TMP/c" check > "$COLIBRI_TMP/make-check.log" 2>&1
+make -C "$COLIBRI_TMP/c" metal-test > "$COLIBRI_TMP/make-metal-test.log" 2>&1
+```
+
+Recorded toolchain and relevant environment:
+
+```text
+date=2026-07-22T05:03:16Z
+uname=Darwin Joshs-MacBook-Pro-2.local 25.5.0 Darwin Kernel Version 25.5.0: Tue Jun  9 22:18:58 PDT 2026; root:xnu-12377.121.10~1/RELEASE_ARM64_T6000 arm64
+macos_product=macOS 26.5.2 build 25F84
+chip=Apple M1 Max
+memory_bytes=34359738368
+arch=arm64
+xcode=Xcode 26.6;Build version 17F113
+clang_path=/usr/bin/clang
+clang_version=Apple clang version 21.0.0 (clang-2100.1.1.101)
+clang_version=Target: arm64-apple-darwin25.5.0
+clang_version=Thread model: posix
+clang_version=InstalledDir: /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin
+clangxx_path=/usr/bin/clang++
+clangxx_version=Apple clang version 21.0.0 (clang-2100.1.1.101)
+clangxx_version=Target: arm64-apple-darwin25.5.0
+clangxx_version=Thread model: posix
+clangxx_version=InstalledDir: /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin
+make_path=/usr/bin/make
+make_version=GNU Make 3.81
+make_version=Copyright (C) 2006  Free Software Foundation, Inc.
+make_version=This is free software; see the source for copying conditions.
+make_version=There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A
+make_version=PARTICULAR PURPOSE.
+make_version=
+make_version=This program built for i386-apple-darwin11.3.0
+python_path=/usr/bin/python3
+python_version=Python 3.9.6
+sdk_path=/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk
+relevant_environment:
+oracle_pin=44e489b196c9b7876b3d37a0570ebf1c6f90f54c
+oracle_pre_status=clean
+workload_precheck=user mlx-bun get active; no bun/python training, inference, benchmark, or Metal test process
+extraction=/var/folders/t3/bdj253f52sv60p_9rmzm86tr0000gn/T/colibri-g0-pin.XXXXXX.phRD7CaPU1
+extracted_git_metadata=absent
+makefile_sha256=b14f9deaf67adc6285facff41914b4bc64089316c83711ae7c7e7dbb731c814a
+glm_c_sha256=3be1b4dd663667c8fa2cfbbacdada3e545ff5f924737a0f5d058708d7bc5ad9d
+```
+
+### `make check`
+
+Complete combined stdout/stderr, verbatim:
+
+```text
+/Applications/Xcode.app/Contents/Developer/usr/bin/make clean
+python3 tools/clean.py
+clean: removed 0 files/dirs
+/Applications/Xcode.app/Contents/Developer/usr/bin/make portable
+/Applications/Xcode.app/Contents/Developer/usr/bin/make glm ARCH=
+make[2]: Circular glm <- glm dependency dropped.
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function glm.c   -o glm -lm -L/opt/homebrew/opt/libomp/lib -lomp
+/Applications/Xcode.app/Contents/Developer/usr/bin/make test
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_json.c -o tests/test_json -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_st.c -o tests/test_st -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function -DST_PREAD_CHUNK=7 tests/test_st_pread.c -o tests/test_st_pread -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_tier.c -o tests/test_tier -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_grammar.c -o tests/test_grammar -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_schema_gbnf.c -o tests/test_schema_gbnf -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_decode_batch.c -o tests/test_decode_batch -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_idot.c -o tests/test_idot -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_i4_grouped.c -o tests/test_i4_grouped -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_stops.c -o tests/test_stops -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_topp.c -o tests/test_topp -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_sample_nan.c -o tests/test_sample_nan -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_kv_alloc.c -o tests/test_kv_alloc -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_i4_acc512.c -o tests/test_i4_acc512 -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_compat_direct.c -o tests/test_compat_direct -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_dsa_select.c -o tests/test_dsa_select -lm -L/opt/homebrew/opt/libomp/lib -lomp
+clang -O3 -Xclang -fopenmp -I/opt/homebrew/opt/libomp/include -Wall -Wextra -Wno-unused-parameter -Wno-misleading-indentation -Wno-unused-function tests/test_logit_nan.c -o tests/test_logit_nan -lm -L/opt/homebrew/opt/libomp/lib -lomp
+python3 tools/run_tests.py tests/test_json tests/test_st tests/test_st_pread tests/test_tier tests/test_grammar tests/test_schema_gbnf tests/test_decode_batch tests/test_idot tests/test_i4_grouped tests/test_stops tests/test_topp tests/test_sample_nan tests/test_kv_alloc tests/test_i4_acc512 tests/test_compat_direct tests/test_dsa_select tests/test_logit_nan
+json tests: ok
+safetensors primitive tests: ok
+test_st_pread: chunk loop + honest truncation error: ok
+tier tests: ok
+test_grammar: ok
+test_schema_gbnf: OK
+decode batch helper tests: ok
+idot kernel exactness (neon): ok
+idot driver exactness (neon): ok
+test_i4_grouped: matmul_i4_grouped vs plain-C dequant reference
+  gs=64, I multiple of gs                    ok (S=2 I=512 O=8 gs=64 ng=8, worst rel 3.1e-08)
+  gs=64, single row single token             ok (S=1 I=128 O=1 gs=64 ng=2, worst rel 9.6e-09)
+  gs=64, nibble edges (0x00/0xFF)            ok (S=1 I=256 O=4 gs=64 ng=4, worst rel 2.6e-08)
+  gs=64, partial last group (I=200)          ok (S=2 I=200 O=4 gs=64 ng=4, worst rel 2e-08)
+  gs=64, I just over a group (I=65)          ok (S=1 I=65 O=3 gs=64 ng=2, worst rel 2e-08)
+  gs=64, I one under a group (I=63)          ok (S=1 I=63 O=3 gs=64 ng=1, worst rel 2.4e-08)
+  gs=64, odd I (I=201)                       ok (S=2 I=201 O=4 gs=64 ng=4, worst rel 1.5e-08)
+  gs=16, odd I (I=33)                        ok (S=1 I=33 O=2 gs=16 ng=3, worst rel 2.2e-08)
+  gs=128 > I=64 (single group)               ok (S=1 I=64 O=4 gs=128 ng=1, worst rel 4.5e-08)
+  gs=128, I multiple of gs                   ok (S=2 I=512 O=4 gs=128 ng=4, worst rel 3.7e-08)
+  gs=64, batch S=8                           ok (S=8 I=320 O=6 gs=64 ng=5, worst rel 7.9e-08)
+test_i4_grouped: ok
+[stop] 5 stop tokens: 100 101 102 103 104 (4 from the tokenizer's special set)
+[stop] 1 stop tokens: 100
+test_stops: stop arming vs incomplete checkpoint metadata
+  tokenizer: special flag parsed, <think>/<tool_call> excluded   ok
+  config eos=[100] + generation_config=[100,101,102] -> 3 stops   ok
+  generation_config.json absent -> config alone, no crash   ok
+  both configs mutilated -> tokenizer still stops all 5 control tokens   ok
+  T=NULL -> config stops only (validation path untouched)   ok
+test_stops: ok
+  ok [V=1 nuc=0.001 shape=uniform keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.500 shape=uniform keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.900 shape=uniform keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.999 shape=uniform keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.001 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.500 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.900 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.999 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.001 shape=geometric keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.500 shape=geometric keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.900 shape=geometric keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.999 shape=geometric keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.001 shape=plateau keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.500 shape=plateau keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.900 shape=plateau keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.999 shape=plateau keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.001 shape=sharptail keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.500 shape=sharptail keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.900 shape=sharptail keep=1 sum=1.0000000000]
+  ok [V=1 nuc=0.999 shape=sharptail keep=1 sum=1.0000000000]
+  ok [V=2 nuc=0.001 shape=uniform keep=1 (ties) sum=1.0000000000]
+  ok [V=2 nuc=0.500 shape=uniform keep=1 (ties) sum=1.0000000000]
+  ok [V=2 nuc=0.900 shape=uniform keep=2 (ties) sum=1.0000000000]
+  ok [V=2 nuc=0.999 shape=uniform keep=2 (ties) sum=1.0000000000]
+  ok [V=2 nuc=0.001 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=2 nuc=0.500 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=2 nuc=0.900 shape=peaked keep=2 sum=0.9999999702]
+  ok [V=2 nuc=0.999 shape=peaked keep=2 sum=0.9999999702]
+  ok [V=2 nuc=0.001 shape=geometric keep=1 sum=1.0000000000]
+  ok [V=2 nuc=0.500 shape=geometric keep=1 sum=1.0000000000]
+  ok [V=2 nuc=0.900 shape=geometric keep=2 sum=1.0000000000]
+  ok [V=2 nuc=0.999 shape=geometric keep=2 sum=1.0000000000]
+  ok [V=2 nuc=0.001 shape=plateau keep=1 (ties) sum=1.0000000000]
+  ok [V=2 nuc=0.500 shape=plateau keep=1 (ties) sum=1.0000000000]
+  ok [V=2 nuc=0.900 shape=plateau keep=2 (ties) sum=1.0000000000]
+  ok [V=2 nuc=0.999 shape=plateau keep=2 (ties) sum=1.0000000000]
+  ok [V=2 nuc=0.001 shape=sharptail keep=1 sum=1.0000000000]
+  ok [V=2 nuc=0.500 shape=sharptail keep=1 sum=1.0000000000]
+  ok [V=2 nuc=0.900 shape=sharptail keep=2 sum=1.0000000000]
+  ok [V=2 nuc=0.999 shape=sharptail keep=2 sum=1.0000000000]
+  ok [V=8 nuc=0.001 shape=uniform keep=1 (ties) sum=1.0000000000]
+  ok [V=8 nuc=0.500 shape=uniform keep=4 (ties) sum=1.0000000000]
+  ok [V=8 nuc=0.900 shape=uniform keep=8 (ties) sum=1.0000000000]
+  ok [V=8 nuc=0.999 shape=uniform keep=8 (ties) sum=1.0000000000]
+  ok [V=8 nuc=0.001 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=8 nuc=0.500 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=8 nuc=0.900 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=8 nuc=0.999 shape=peaked keep=5 sum=0.9999999900]
+  ok [V=8 nuc=0.001 shape=geometric keep=1 sum=1.0000000000]
+  ok [V=8 nuc=0.500 shape=geometric keep=4 sum=1.0000000447]
+  ok [V=8 nuc=0.900 shape=geometric keep=8 sum=1.0000000000]
+  ok [V=8 nuc=0.999 shape=geometric keep=8 sum=1.0000000000]
+  ok [V=8 nuc=0.001 shape=plateau keep=1 (ties) sum=1.0000000000]
+  ok [V=8 nuc=0.500 shape=plateau keep=4 (ties) sum=1.0000000000]
+  ok [V=8 nuc=0.900 shape=plateau keep=7 (ties) sum=1.0000000447]
+  ok [V=8 nuc=0.999 shape=plateau keep=8 (ties) sum=1.0000000075]
+  ok [V=8 nuc=0.001 shape=sharptail keep=1 sum=1.0000000000]
+  ok [V=8 nuc=0.500 shape=sharptail keep=3 sum=1.0000000149]
+  ok [V=8 nuc=0.900 shape=sharptail keep=7 sum=0.9999999776]
+  ok [V=8 nuc=0.999 shape=sharptail keep=8 sum=1.0000000075]
+  ok [V=64 nuc=0.001 shape=uniform keep=1 (ties) sum=1.0000000000]
+  ok [V=64 nuc=0.500 shape=uniform keep=32 (ties) sum=1.0000000000]
+  ok [V=64 nuc=0.900 shape=uniform keep=58 (ties) sum=0.9999999963]
+  ok [V=64 nuc=0.999 shape=uniform keep=64 (ties) sum=1.0000000000]
+  ok [V=64 nuc=0.001 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=64 nuc=0.500 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=64 nuc=0.900 shape=peaked keep=2 sum=1.0000000186]
+  ok [V=64 nuc=0.999 shape=peaked keep=39 sum=0.9999999860]
+  ok [V=64 nuc=0.001 shape=geometric keep=1 sum=1.0000000000]
+  ok [V=64 nuc=0.500 shape=geometric keep=32 sum=1.0000000149]
+  ok [V=64 nuc=0.900 shape=geometric keep=58 sum=0.9999999814]
+  ok [V=64 nuc=0.999 shape=geometric keep=64 sum=0.9999999879]
+  ok [V=64 nuc=0.001 shape=plateau keep=1 (ties) sum=1.0000000000]
+  ok [V=64 nuc=0.500 shape=plateau keep=5 (ties) sum=1.0000000149]
+  ok [V=64 nuc=0.900 shape=plateau keep=13 (ties) sum=0.9999999702]
+  ok [V=64 nuc=0.999 shape=plateau keep=35 (ties) sum=0.9999999879]
+  ok [V=64 nuc=0.001 shape=sharptail keep=1 (ties) sum=1.0000000000]
+  ok [V=64 nuc=0.500 shape=sharptail keep=4 (ties) sum=1.0000000596]
+  ok [V=64 nuc=0.900 shape=sharptail keep=10 (ties) sum=0.9999999925]
+  ok [V=64 nuc=0.999 shape=sharptail keep=12 (ties) sum=0.9999999925]
+  ok [V=257 nuc=0.001 shape=uniform keep=1 (ties) sum=1.0000000000]
+  ok [V=257 nuc=0.500 shape=uniform keep=129 (ties) sum=1.0000000563]
+  ok [V=257 nuc=0.900 shape=uniform keep=232 (ties) sum=0.9999999963]
+  ok [V=257 nuc=0.999 shape=uniform keep=257 (ties) sum=0.9999999998]
+  ok [V=257 nuc=0.001 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=257 nuc=0.500 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=257 nuc=0.900 shape=peaked keep=11 sum=0.9999999867]
+  ok [V=257 nuc=0.999 shape=peaked keep=203 sum=0.9999999721]
+  ok [V=257 nuc=0.001 shape=geometric keep=1 sum=1.0000000000]
+  ok [V=257 nuc=0.500 shape=geometric keep=117 sum=0.9999999702]
+  ok [V=257 nuc=0.900 shape=geometric keep=227 sum=0.9999999774]
+  ok [V=257 nuc=0.999 shape=geometric keep=257 sum=0.9999999749]
+  ok [V=257 nuc=0.001 shape=plateau keep=1 (ties) sum=1.0000000000]
+  ok [V=257 nuc=0.500 shape=plateau keep=5 (ties) sum=0.9999999404]
+  ok [V=257 nuc=0.900 shape=plateau keep=13 (ties) sum=0.9999999590]
+  ok [V=257 nuc=0.999 shape=plateau keep=35 (ties) sum=0.9999999854]
+  ok [V=257 nuc=0.001 shape=sharptail keep=1 (ties) sum=1.0000000000]
+  ok [V=257 nuc=0.500 shape=sharptail keep=5 (ties) sum=0.9999999553]
+  ok [V=257 nuc=0.900 shape=sharptail keep=10 (ties) sum=0.9999999851]
+  ok [V=257 nuc=0.999 shape=sharptail keep=12 (ties) sum=0.9999999925]
+  ok [V=1519 nuc=0.001 shape=uniform keep=2 (ties) sum=1.0000000000]
+  ok [V=1519 nuc=0.500 shape=uniform keep=760 (ties) sum=1.0000000428]
+  ok [V=1519 nuc=0.900 shape=uniform keep=1368 (ties) sum=1.0000000075]
+  ok [V=1519 nuc=0.999 shape=uniform keep=1518 (ties) sum=1.0000000106]
+  ok [V=1519 nuc=0.001 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=1519 nuc=0.500 shape=peaked keep=1 sum=1.0000000000]
+  ok [V=1519 nuc=0.900 shape=peaked keep=404 sum=0.9999999575]
+  ok [V=1519 nuc=0.999 shape=peaked keep=1380 sum=1.0000000113]
+  ok [V=1519 nuc=0.001 shape=geometric keep=1 sum=1.0000000000]
+  ok [V=1519 nuc=0.500 shape=geometric keep=410 sum=1.0000000135]
+  ok [V=1519 nuc=0.900 shape=geometric keep=1118 sum=0.9999999862]
+  ok [V=1519 nuc=0.999 shape=geometric keep=1514 sum=1.0000000146]
+  ok [V=1519 nuc=0.001 shape=plateau keep=1 (ties) sum=1.0000000000]
+  ok [V=1519 nuc=0.500 shape=plateau keep=5 (ties) sum=0.9999999404]
+  ok [V=1519 nuc=0.900 shape=plateau keep=13 (ties) sum=0.9999999590]
+  ok [V=1519 nuc=0.999 shape=plateau keep=35 (ties) sum=0.9999999854]
+  ok [V=1519 nuc=0.001 shape=sharptail keep=1 (ties) sum=1.0000000000]
+  ok [V=1519 nuc=0.500 shape=sharptail keep=5 (ties) sum=1.0000000149]
+  ok [V=1519 nuc=0.900 shape=sharptail keep=10 (ties) sum=0.9999999702]
+  ok [V=1519 nuc=0.999 shape=sharptail keep=12 (ties) sum=1.0000000186]
+  ok [guard-off nuc=1.0 keep=256]
+  ok [guard-off nuc=0.0 keep=256]
+  ok [V=1 keep=1]
+
+test_topp: 123 cases run, 0 failure(s)
+test_topp: ok
+[SAMPLE] warning: non-finite logits (NaN/Inf) — falling back to argmax; output may be degraded. This usually means a numerical blow-up upstream.
+  logit sani: distribuzione valida, picco vivo            ok
+  NaN/+Inf iniettato: argmax dei finiti vince, mai 0/NaN   ok
+  tutti non-finiti: nessun crash, buffer valido               ok
+test_sample_nan: ok
+OK kv_alloc re-allocation
+test_i4_acc512: skipped (no AVX-512 on this build)
+compat direct tests: skipped (POSIX has native O_DIRECT)
+  ok [nk=1 keep=1 shape=random]
+  ok [nk=1 keep=1 shape=peaked]
+  ok [nk=1 keep=1 shape=decreasing]
+  ok [nk=1 keep=1 shape=increasing]
+  ok [nk=1 keep=1 shape=plateau]
+  ok [nk=1 keep=1 shape=all-equal]
+  ok [nk=2 keep=1 shape=random]
+  ok [nk=2 keep=1 shape=peaked]
+  ok [nk=2 keep=1 shape=decreasing]
+  ok [nk=2 keep=1 shape=increasing]
+  ok [nk=2 keep=1 shape=plateau]
+  ok [nk=2 keep=1 shape=all-equal]
+  ok [nk=8 keep=1 shape=random]
+  ok [nk=8 keep=8 shape=random]
+  ok [nk=8 keep=1 shape=peaked]
+  ok [nk=8 keep=8 shape=peaked]
+  ok [nk=8 keep=1 shape=decreasing]
+  ok [nk=8 keep=8 shape=decreasing]
+  ok [nk=8 keep=1 shape=increasing]
+  ok [nk=8 keep=8 shape=increasing]
+  ok [nk=8 keep=1 shape=plateau]
+  ok [nk=8 keep=8 shape=plateau]
+  ok [nk=8 keep=1 shape=all-equal]
+  ok [nk=8 keep=8 shape=all-equal]
+  ok [nk=64 keep=1 shape=random]
+  ok [nk=64 keep=8 shape=random]
+  ok [nk=64 keep=1 shape=peaked]
+  ok [nk=64 keep=8 shape=peaked]
+  ok [nk=64 keep=1 shape=decreasing]
+  ok [nk=64 keep=8 shape=decreasing]
+  ok [nk=64 keep=1 shape=increasing]
+  ok [nk=64 keep=8 shape=increasing]
+  ok [nk=64 keep=1 shape=plateau]
+  ok [nk=64 keep=8 shape=plateau]
+  ok [nk=64 keep=1 shape=all-equal]
+  ok [nk=64 keep=8 shape=all-equal]
+  ok [nk=2049 keep=1 shape=random]
+  ok [nk=2049 keep=8 shape=random]
+  ok [nk=2049 keep=256 shape=random]
+  ok [nk=2049 keep=1024 shape=random]
+  ok [nk=2049 keep=2048 shape=random]
+  ok [nk=2049 keep=1 shape=peaked]
+  ok [nk=2049 keep=8 shape=peaked]
+  ok [nk=2049 keep=256 shape=peaked]
+  ok [nk=2049 keep=1024 shape=peaked]
+  ok [nk=2049 keep=2048 shape=peaked]
+  ok [nk=2049 keep=1 shape=decreasing]
+  ok [nk=2049 keep=8 shape=decreasing]
+  ok [nk=2049 keep=256 shape=decreasing]
+  ok [nk=2049 keep=1024 shape=decreasing]
+  ok [nk=2049 keep=2048 shape=decreasing]
+  ok [nk=2049 keep=1 shape=increasing]
+  ok [nk=2049 keep=8 shape=increasing]
+  ok [nk=2049 keep=256 shape=increasing]
+  ok [nk=2049 keep=1024 shape=increasing]
+  ok [nk=2049 keep=2048 shape=increasing]
+  ok [nk=2049 keep=1 shape=plateau]
+  ok [nk=2049 keep=8 shape=plateau]
+  ok [nk=2049 keep=256 shape=plateau]
+  ok [nk=2049 keep=1024 shape=plateau]
+  ok [nk=2049 keep=2048 shape=plateau]
+  ok [nk=2049 keep=1 shape=all-equal]
+  ok [nk=2049 keep=8 shape=all-equal]
+  ok [nk=2049 keep=256 shape=all-equal]
+  ok [nk=2049 keep=1024 shape=all-equal]
+  ok [nk=2049 keep=2048 shape=all-equal]
+  ok [nk=4097 keep=1 shape=random]
+  ok [nk=4097 keep=8 shape=random]
+  ok [nk=4097 keep=256 shape=random]
+  ok [nk=4097 keep=1024 shape=random]
+  ok [nk=4097 keep=2048 shape=random]
+  ok [nk=4097 keep=1 shape=peaked]
+  ok [nk=4097 keep=8 shape=peaked]
+  ok [nk=4097 keep=256 shape=peaked]
+  ok [nk=4097 keep=1024 shape=peaked]
+  ok [nk=4097 keep=2048 shape=peaked]
+  ok [nk=4097 keep=1 shape=decreasing]
+  ok [nk=4097 keep=8 shape=decreasing]
+  ok [nk=4097 keep=256 shape=decreasing]
+  ok [nk=4097 keep=1024 shape=decreasing]
+  ok [nk=4097 keep=2048 shape=decreasing]
+  ok [nk=4097 keep=1 shape=increasing]
+  ok [nk=4097 keep=8 shape=increasing]
+  ok [nk=4097 keep=256 shape=increasing]
+  ok [nk=4097 keep=1024 shape=increasing]
+  ok [nk=4097 keep=2048 shape=increasing]
+  ok [nk=4097 keep=1 shape=plateau]
+  ok [nk=4097 keep=8 shape=plateau]
+  ok [nk=4097 keep=256 shape=plateau]
+  ok [nk=4097 keep=1024 shape=plateau]
+  ok [nk=4097 keep=2048 shape=plateau]
+  ok [nk=4097 keep=1 shape=all-equal]
+  ok [nk=4097 keep=8 shape=all-equal]
+  ok [nk=4097 keep=256 shape=all-equal]
+  ok [nk=4097 keep=1024 shape=all-equal]
+  ok [nk=4097 keep=2048 shape=all-equal]
+  ok [nk=8193 keep=1 shape=random]
+  ok [nk=8193 keep=8 shape=random]
+  ok [nk=8193 keep=256 shape=random]
+  ok [nk=8193 keep=1024 shape=random]
+  ok [nk=8193 keep=2048 shape=random]
+  ok [nk=8193 keep=1 shape=peaked]
+  ok [nk=8193 keep=8 shape=peaked]
+  ok [nk=8193 keep=256 shape=peaked]
+  ok [nk=8193 keep=1024 shape=peaked]
+  ok [nk=8193 keep=2048 shape=peaked]
+  ok [nk=8193 keep=1 shape=decreasing]
+  ok [nk=8193 keep=8 shape=decreasing]
+  ok [nk=8193 keep=256 shape=decreasing]
+  ok [nk=8193 keep=1024 shape=decreasing]
+  ok [nk=8193 keep=2048 shape=decreasing]
+  ok [nk=8193 keep=1 shape=increasing]
+  ok [nk=8193 keep=8 shape=increasing]
+  ok [nk=8193 keep=256 shape=increasing]
+  ok [nk=8193 keep=1024 shape=increasing]
+  ok [nk=8193 keep=2048 shape=increasing]
+  ok [nk=8193 keep=1 shape=plateau]
+  ok [nk=8193 keep=8 shape=plateau]
+  ok [nk=8193 keep=256 shape=plateau]
+  ok [nk=8193 keep=1024 shape=plateau]
+  ok [nk=8193 keep=2048 shape=plateau]
+  ok [nk=8193 keep=1 shape=all-equal]
+  ok [nk=8193 keep=8 shape=all-equal]
+  ok [nk=8193 keep=256 shape=all-equal]
+  ok [nk=8193 keep=1024 shape=all-equal]
+  ok [nk=8193 keep=2048 shape=all-equal]
+  ok [keep==nk nk=100]
+  ok [keep==1 argmax=19]
+  ok [all-equal keep=500 -> positions 0..499]
+
+test_dsa_select: 129 cases run, 0 failure(s)
+test_dsa_select: ok
+[SAMPLE] warning: non-finite logits (NaN/Inf) — falling back to argmax; output may be degraded. This usually means a numerical blow-up upstream.
+OK test_logit_nan: argmax_v NaN-skip + dist_build finite-collapse
+python3 -m unittest discover -s tests -p 'test_*.py'
+......ss.sss........................[api] 127.0.0.1 - "OPTIONS /v1/chat/completions HTTP/1.1" 204 -
+.[api] 127.0.0.1 - "POST /v1/chat/completions HTTP/1.1" 200 -
+.[api] 127.0.0.1 - "GET /health HTTP/1.1" 200 -
+.[api] 127.0.0.1 - "POST /v1/completions HTTP/1.1" 200 -
+.[api] 127.0.0.1 - "GET /v1/models HTTP/1.1" 200 -
+[api] 127.0.0.1 - "GET /v1/models HTTP/1.1" 401 -
+.[api] 127.0.0.1 - "GET /profile HTTP/1.1" 200 -
+[api] 127.0.0.1 - "GET /profile HTTP/1.1" 200 -
+.[api] 127.0.0.1 - "POST /v1/completions HTTP/1.1" 400 -
+.[api] 127.0.0.1 - "POST /v1/chat/completions HTTP/1.1" 400 -
+.[api] 127.0.0.1 - "POST /v1/chat/completions HTTP/1.1" 400 -
+.[api] 127.0.0.1 - "POST /v1/chat/completions HTTP/1.1" 200 -
+....[api] 127.0.0.1 - "POST /v1/chat/completions HTTP/1.1" 429 -
+[api] 127.0.0.1 - "POST /v1/chat/completions HTTP/1.1" 200 -
+......[api] 127.0.0.1 - "GET / HTTP/1.1" 200 -
+[api] 127.0.0.1 - "GET /%2e%2e/dist-private/secret.txt HTTP/1.1" 404 -
+.....[api] tool-calls: 1 total, 1 strict, 0 de-mangled [CLEAN]
+.[api] tool-calls: 1 total, 1 strict, 0 de-mangled [CLEAN]
+.[api] tool-calls: 1 total, 1 strict, 0 de-mangled [CLEAN]
+.......................
+----------------------------------------------------------------------
+Ran 84 tests in 5.427s
+
+OK (skipped=5)
+```
+
+Exit status: `0`.
+
+### `make metal-test`
+
+A second workload check immediately before this command found the same
+user-owned download and no user training, inference, benchmark, Colibri test,
+or other Bun/Python MLX compute process.
+
+Complete combined stdout/stderr, verbatim:
+
+```text
+clang++ -x objective-c++ -std=gnu++17 -fobjc-arc -O3 tests/test_backend_metal.mm backend_metal.mm -framework Metal -framework Foundation -o backend_metal_test
+./backend_metal_test
+Metal backend kernel tests:
+  int8 gate/up S=1       nerr=2.36e-06  ok
+  int4 gate/up S=1       nerr=1.98e-06  ok
+  int4 down S=1          nerr=1.03e-06  ok
+  int2 gate/up S=1       nerr=1.72e-06  ok
+  f32  S=1               nerr=1.77e-06  ok
+  int8 gate/up S=4       nerr=2.36e-06  ok
+  int4 gate/up S=7 (odd) nerr=1.98e-06  ok
+  int4 non-mult-4 dims   nerr=3.78e-06  ok
+Metal batched moe_block tests:
+  moe decode nb=8        R=8 nerr=2.53e-06  ok
+  moe ragged nb=6        R=16 nerr=2.45e-06  ok
+Metal large-batch gemm test:
+  gemm S=64 int4          nerr=2.85e-06  ok
+Metal fused attention tests:
+  attn S=1 pos=0           nerr=4.04e-06 cache=1.38e-05  ok
+  attn S=1 pos=37          nerr=4.19e-06 cache=1.24e-05  ok
+  attn S=4 pos=12 (MTP)    nerr=5.33e-06 cache=1.34e-05  ok
+  attn S=3 pos=0           nerr=5.30e-06 cache=1.31e-05  ok
+metal backend tests: ok
+```
+
+Exit status: `0`.
+
+### Oracle postcondition
+
+At `2026-07-22T05:06:07Z`, after both suites completed:
+
+```text
+HEAD=44e489b196c9b7876b3d37a0570ebf1c6f90f54c
+status --short: <empty>
+git diff --quiet exit status: 0
+git diff --cached --quiet exit status: 0
+```
+
+No build, test, or mutating command was run from the oracle checkout, no
+generated artifact was copied back, and no file in that checkout was changed.
+
+## Target-machine disk preflight — in-progress snapshot
+
+This is a **moving snapshot, not the final stable G0 disk baseline**. A
+user-owned artifact download was already active throughout the measurement:
+
+```text
+46623 Tue Jul 21 23:00:44 2026           08:37 bun /Users/joshrossi/.bun/bin/mlx-bun get mateogrgic/GLM-5.2-colibri-int4-with-int8-mtp
+```
+
+The process was observed only. It was not stopped, signalled, reprioritized, or
+otherwise altered. No snapshot was deleted and no space was purged. Capacity
+therefore continued to change while these commands ran; repeat this preflight
+after the download settles and before recording the full-model baseline.
+
+### Timestamp and target identity
+
+Command:
+
+```sh
+date '+%Y-%m-%dT%H:%M:%S%z (%Z)'
+system_profiler SPHardwareDataType 2>/dev/null | awk '/Model Name:|Model Identifier:|Chip:|Total Number of Cores:|Memory:/{print}'
+sw_vers
+sysctl -n hw.memsize
+readlink /etc/localtime
+mount | awk '$3 == "/" {print}'
+```
+
+Output:
+
+```text
+2026-07-21T23:09:50-0600 (CST)
+      Model Name: MacBook Pro
+      Model Identifier: MacBookPro18,2
+      Chip: Apple M1 Max
+      Total Number of Cores: 10 (8 Performance and 2 Efficiency)
+      Memory: 32 GB
+ProductName:		macOS
+ProductVersion:		26.5.2
+BuildVersion:		25F84
+34359738368
+/var/db/timezone/zoneinfo/America/El_Salvador
+/dev/disk3s3s1 on / (apfs, sealed, local, read-only, journaled)
+```
+
+The qualifying machine is therefore the plan's M1 Max MacBook Pro with 32 GiB
+physical memory, running macOS 26.5.2 in the America/El_Salvador timezone. The
+root is the sealed APFS system snapshot; the writable artifact data resides in
+the same APFS container through `/System/Volumes/Data`.
+
+### Filesystem capacity
+
+Measurement window: `2026-07-21T23:09:20-0600` through
+`2026-07-21T23:09:21-0600` (CST, America/El_Salvador).
+
+Commands:
+
+```sh
+df -h /
+df -k /
+df -h /System/Volumes/Data
+df -k /System/Volumes/Data
+```
+
+Output:
+
+```text
+Filesystem        Size    Used   Avail Capacity iused ifree %iused  Mounted on
+/dev/disk3s3s1   926Gi    12Gi   506Gi     3%    459k  4.3G    0%   /
+Filesystem     1024-blocks      Used Available Capacity iused      ifree %iused  Mounted on
+/dev/disk3s3s1   971350180  12270752 531096252     3%  458726 4293644559    0%   /
+Filesystem      Size    Used   Avail Capacity iused ifree %iused  Mounted on
+/dev/disk3s1   926Gi   398Gi   506Gi    45%    4.5M  5.3G    0%   /System/Volumes/Data
+Filesystem   1024-blocks      Used Available Capacity iused      ifree %iused  Mounted on
+/dev/disk3s1   971350180 417631284 531096252    45% 4512220 5310962520    0%   /System/Volumes/Data
+```
+
+The System and Data volumes share the same 506 GiB available APFS-container
+capacity; their different `Used` columns reflect volume roles and are not
+additive.
+
+Relevant `diskutil apfs list` command and filtered output:
+
+```sh
+diskutil apfs list | awk '/Container disk3 /{show=1} show && /APFS Container Reference:|Size \\(Capacity Ceiling\\):|Capacity In Use By Volumes:|Capacity Not Allocated:|Physical Store Disk:|Mount Point:|Capacity Consumed:/{print}'
+```
+
+```text
+    APFS Container Reference:     disk3
+    Size (Capacity Ceiling):      994662584320 B (994.7 GB)
+    Capacity In Use By Volumes:   450821074944 B (450.8 GB) (45.3% used)
+    Capacity Not Allocated:       543841509376 B (543.8 GB) (54.7% free)
+    |   APFS Physical Store Disk:   disk0s2
+    |   Mount Point:               /System/Volumes/Data
+    |   Capacity Consumed:         427655487488 B (427.7 GB)
+    |   Mount Point:               Not Mounted
+    |   Capacity Consumed:         12565250048 B (12.6 GB)
+    |   Snapshot Mount Point:      /
+    |   Mount Point:               /System/Volumes/Preboot
+    |   Capacity Consumed:         9037856768 B (9.0 GB)
+    |   Mount Point:               Not Mounted
+    |   Capacity Consumed:         1308078080 B (1.3 GB)
+        Mount Point:               /System/Volumes/VM
+        Capacity Consumed:         20480 B (20.5 KB)
+```
+
+Relevant writable-volume identity from `diskutil info
+/System/Volumes/Data`:
+
+```text
+   Device Identifier:         disk3s1
+   Volume Name:               Macintosh HD - Data
+   Mount Point:               /System/Volumes/Data
+   File System Personality:   APFS
+   Container Total Space:     994.7 GB (994662584320 Bytes) (exactly 1942700360 512-Byte-Units)
+   Container Free Space:      543.8 GB (543841349632 Bytes) (exactly 1062190136 512-Byte-Units)
+   Device Location:           Internal
+   Solid State:               Yes
+```
+
+### Purgeable and snapshot evidence
+
+Foundation's read-only APFS capacity keys were queried because `df` and
+`diskutil apfs list` expose hard unallocated capacity but do not report one
+stable scalar named “purgeable”:
+
+```sh
+osascript -l JavaScript -e 'ObjC.import("Foundation"); const u=$.NSURL.fileURLWithPath("/"); function g(k){const v=Ref(); const e=Ref(); if(!u.getResourceValueForKeyError(v,k,e)) return "error"; return ObjC.unwrap(v[0]);} JSON.stringify({available:g($.NSURLVolumeAvailableCapacityKey),important:g($.NSURLVolumeAvailableCapacityForImportantUsageKey),opportunistic:g($.NSURLVolumeAvailableCapacityForOpportunisticUsageKey),total:g($.NSURLVolumeTotalCapacityKey)});'
+```
+
+```json
+{"available":543824449536,"important":590191163121,"opportunistic":643859215665,"total":994662584320}
+```
+
+`available` agrees with the hard `diskutil` value to normal sampling drift.
+The `important` estimate was 46,366,713,585 bytes above hard available at this
+instant, but these Foundation estimates are policy-dependent advisory reclaim
+figures—not guaranteed immediately purgeable bytes—and moved while the download
+was active. They are not needed to pass the hard-space threshold.
+
+Read-only snapshot commands:
+
+```sh
+diskutil apfs listSnapshots /
+tmutil listlocalsnapshots /
+```
+
+Output:
+
+```text
+Snapshot for disk3s3s1 (1 found)
+|
++-- BE13DEB1-389A-4346-A93F-502BBAF6176B
+    Name:        com.apple.os.update-5B92CE4BA1034457A0921532291F4A4AD939CBE48D3A1381163A9AD3687D5694
+    XID:         5287272
+    Purgeable:   No
+    NOTE:        This snapshot limits the minimum size of APFS Container disk3
+Snapshots for disk /:
+```
+
+There were no Time Machine local snapshots listed. The sole root-volume OS
+update snapshot is explicitly non-purgeable.
+
+### `>=400 GB` verdict
+
+**PASS for this in-progress snapshot:** APFS reported
+`543,841,509,376` bytes (`543.8 GB`, approximately `506 GiB`) hard unallocated,
+which is `143.8 GB` above the plan's `>=400 GB` pre-download/pre-conversion
+threshold without counting advisory reclaimable capacity.
+
+This is deliberately not treated as the final stable G0 baseline. The active
+approximately 372 GB artifact transfer can materially change the value. Re-run
+the same commands after that user-owned transfer completes and before any
+conversion or full-model oracle measurement.
+
+## Stable post-download machine and build preflight (2026-07-22)
+
+The post-restart target was an Apple M1 Max with 10 cores and 32 GiB unified
+memory, running macOS 26.5.2 (`25F84`, Darwin 25.5.0). The model is on the
+internal APFS SSD (`Apple Fabric`). During the stable audit APFS reported about
+`174.1 GB` hard-free, with no local Time Machine snapshots. Before inference,
+`memory_pressure -Q` reported 93% free, swap and compressed pages were zero,
+and no Colibri, MLX, Bun, Python model, or Hugging Face transfer process was
+active. The 400 GB value above was the satisfied pre-download floor; it is not
+a post-download runtime requirement.
+
+The oracle source checkout was returned to a clean, build-artifact-free
+`44e489b196c9b7876b3d37a0570ebf1c6f90f54c`. A `git archive` of that exact pin
+was built separately under ignored `runs/colibri-g0/` with:
+
+```bash
+make -C <exact-pin-archive>/c glm METAL=1
+make -C <exact-pin-archive>/c metal-test
+```
+
+Apple clang 21.0.0 built the engine with `-O3`, Homebrew OpenMP, and
+`-DCOLI_METAL`. Every standalone Metal int8/int4/int2 matmul, batched MoE,
+large GEMM, and fused-attention comparison passed. The tested engine SHA-256
+is `fa2b00cb7b8b5fbfb0c43908567c07fdd2bb9024ec47502697cb01885ca2b501`.
+
+Colibri's Python doctor correctly analyzes 144 shards, 19,456 experts,
+`10,877,286,144` dense bytes, `372,848,459,776` expert bytes, and
+`18,915,328` bytes per typical expert. Its conservative 4K-context plan did
+not pass on the then-current reclaimable-memory snapshot: auto sizing produced
+zero cache slots and a forced 25 GB budget exceeded reported availability.
+Separately, the C profiler's `hw_probe` has no macOS branch and prints
+`unknown CPU | RAM 0.0 GB total, 0.0 GB available`; the engine's distinct
+`host_statistics64` safety path does have macOS support. This is an upstream
+telemetry bug, not evidence that unified memory was unavailable.
+
+## Bounded full-model proof run (2026-07-22)
+
+At the user's explicit direction, the exact-pin Metal engine was run against
+the complete 357.404 GiB snapshot with a deliberately bounded configuration:
+18 GB RAM budget, 128-token context, 19-token deterministic prompt, at most 8
+generated tokens, true top-8 expert routing, no cache-aware routing, expert
+budget, pilot, repin, or autopin. Both cells used `COLI_METAL=1`, `DIRECT=1`,
+`PIPE=1`, six loader workers, `TEMP=0`, `SEED=1`, `PROF=1`, and int4 main
+expert/dense interpretation. MTP-off set `MTP=0 DRAFT=0`; MTP-on set
+`MTP=1 DRAFT=3 SPEC_PIN=1`. DSA was off because the artifact has no indexer
+weights.
+
+Both cells emitted the same six token IDs (`17 10 17 28 19 13`) and the same
+correct text: `2+2=4.`
+
+| Metric | MTP off | MTP on |
+|---|---:|---:|
+| Engine load | 3.22 s | 2.44 s |
+| Prefill, 19 tokens | 20.94 s | 21.04 s |
+| Decode, 6 tokens | 17.61 s | 14.44 s |
+| Decode rate | 0.34 tok/s | 0.42 tok/s |
+| Process wall time | 42.62 s | 38.36 s |
+| Tokens/forward | 1.00 | 3.00 |
+| MTP accepted/proposed | 0/0 | 4/6 (67%) |
+| LRU hit rate | 4.8% | 1.5% |
+| Expert cache cap | 2/layer | 1/layer |
+| Expert payload fetched | 64.823 GB | 65.576 GB |
+| Read service / felt wait | 52.351 / 9.929 s | 56.293 / 9.696 s |
+| Maximum RSS (`time -l`) | 6.19 GB | 10.56 GB |
+| Peak footprint (`time -l`) | 14.98 GB | 16.29 GB |
+| Swap | 0 | 0 |
+
+For this tiny proof, MTP-on improved reported decode rate by about 23.5%,
+reduced decode wall time by about 18.0%, and reduced whole-process wall time by
+about 10.0%. Those deltas are not benchmark claims: this is one sequential
+pair, only six generated tokens, the cells had different LRU capacities, and
+the MTP files had been read during integrity verification. It establishes the
+important facts that the real GLM-5.2 container loads on the 32 GB M1 Max,
+Metal unified-memory zero-copy executes with no CPU fallback, output is sane,
+the corrected MTP layer activates, SPEC_PIN activates, and neural drafts are
+accepted.
+
+Raw machine-local evidence and SHA-256 values:
+
+- `runs/colibri-g0/results-20260722/mtp-off.log`:
+  `593a9274f70e7b2f80a5f90bc5d3edea948d77199897464c3ec3134b989f974b`
+- `runs/colibri-g0/results-20260722/mtp-on.log`:
+  `58d9715d7d81ef94aeaf9a70e1fc1ca2442341c4dcd9c380000dc7bf4bc1bb85`
+- `mtp-off.stats`:
+  `767abcfd2a21af8797509586b2f0e6ab990fb28ddefd40445ca42cf8d252933d`
+- `mtp-on.stats`:
+  `c409ed39ecb97b6a9da999a121db372fc25db79f8eacffd4d5afcc589bbd59d1`
+
+At the time of this bounded proof, G0 remained open pending the repeated
+matrix, complete main-shard hashes, DSA decision, exact TTFT instrumentation,
+acceptance-length distribution, and numeric oracle. The closure section below
+supersedes this historical state.
+
+## G0 item 4: model-free capture and derived scaffolding (2026-07-21)
+
+An incomplete model-free starting package is checked in at
+[`../../fixtures/colibri-glm52/`](../../fixtures/colibri-glm52/README.md), with
+an explicit exact-pin capture orchestrator at
+[`../../scripts/capture-colibri-glm52-oracle.ts`](../../scripts/capture-colibri-glm52-oracle.ts),
+capture helpers at `scripts/colibri-glm52-capture.{py,c}`, the ordinary
+standalone generator at
+[`../../scripts/gen-colibri-glm52-fixtures.ts`](../../scripts/gen-colibri-glm52-fixtures.ts)
+and consuming model-free tests at
+[`../../tests/colibri-glm52-fixtures.test.ts`](../../tests/colibri-glm52-fixtures.test.ts).
+Git is the integrity root. The manifest records reproducibility hashes but does
+not present itself as an authenticity root.
+
+The capture script first validates that the external checkout has exact HEAD,
+empty porcelain, and clean cached/worktree diffs, then executes only a temporary
+`git archive`. The recorded run used Apple M1 Max ARM64, Apple clang 21.0.0,
+CPython 3.14.5 with existing NumPy 2.4.6, and Bun 1.3.14. It imported the
+archived pinned `quant_int8`/`quant_int4_grouped`, and compiled a temporary
+harness that includes archived `glm.c`/`tier.h`. Direct captured constants are
+limited to Python quantization, ARM `matmul_q`/`matmul_i4_grouped`, DSA
+partial-select/tie scans, LFRU score/admission/tie/wrap/decay, and elementary
+RMSNorm/sigmoid outputs. Exact tree/blob hashes, source SHA-256 values, helper
+hashes, toolchain, target, and executed commands live in
+`oracle-capture.json`. No install, download, server, model load, or external
+checkout mutation occurred.
+
+The tiny SwiGLU/residual spine, top-8 router composition, LRU trace, and greedy
+MTP state trace are explicitly `derived_canonical`, not captured Colibri
+runtime outputs. The MTP scaffold covers partial accept, immediate reject,
+full accept, accepted-only absorption, next-logit/`hlast` row choice, and
+logical main-KV rollback. EOS/special-stop handling, length clamps, adaptive
+guard, grammar precedence, sampling rejection, teacher-forced logits, numeric
+GLM/MLA/MTP-head/KV values, and neural acceptance remain gaps.
+
+Residual blocker: item 4 and G0 remain incomplete until a real numeric model
+oracle is recorded and the required cleared-M1-Max-32-GB same-machine baseline
+is run. The earlier approximately 679 GB preflight remains the recorded stable
+capacity note; the later moving download snapshot does not replace it, and a
+stable refresh after the user-owned transfer settles remains pending.
+
+## G0 closure: full oracle and same-machine baseline (2026-07-22)
+
+This section supersedes the two historical "G0 remains open" statements above.
+The target machine was restarted and had no competing model/download process,
+but the initial proof and three MTP-file hashes occurred before the controlled
+cells. The reproducible classification is therefore:
+
+- **process-cold/direct-I/O:** a new `glm` process, empty explicit LRU,
+  `DIRECT=1` (macOS `F_NOCACHE`), no pin/autopin/repin/pilot/cache-aware route;
+- **session-LRU-warm:** request two in the same PID after `RESET` clears KV and
+  leaves the explicit expert LRU resident;
+- **not boot-cold:** no claim that the filesystem had been untouched since
+  boot. Direct/no-cache expert reads make literal boot cache state irrelevant
+  to the streamed expert payload.
+
+All controlled cells used the same 32-token prompt, `NGEN=64`, `TEMP=0`,
+`SEED=1`, `RAM_GB=18`, `CTX=128`, cap 1/layer, Metal unified-memory zero-copy,
+six direct-I/O workers, true top-8, and int4 main weights. MTP-on changed
+`MTP/DRAFT` from `0/0` to `1/3`; `SPEC_PIN=1` was pinned in both modes and is
+operative only with MTP. DSA stayed off.
+
+### Earlier standalone decode-separated corroboration
+
+This three-repeat harness was the first controlled process-cold result. It is
+retained because it separates prefill from decode, but the final G0 closure
+aggregate is the stricter six-process/twelve-turn matrix in the next section.
+
+| Metric (three repeats; median, range where useful) | MTP off | MTP on |
+|---|---:|---:|
+| Prefill | 30.42 s | 31.56 s |
+| Decode, 64 tokens | 164.08 s (158.40-166.56) | 207.83 s (194.21-210.40) |
+| Decode rate | 0.39 tok/s (0.38-0.40) | 0.31 tok/s (0.30-0.33) |
+| Whole-process wall | 197.57 s | 243.70 s |
+| Main forwards / tokens per forward | 63 / 1.02 | 30 / 2.13 |
+| MTP accepted / proposed | 0/0 | 34/90 every repeat |
+| LRU hit rate | 2.3% | 1.3% |
+| Expert payload fetched | 698.902 GB | 983.671 GB |
+| Peak footprint | 13.60 GB | 16.98 GB |
+| Process swaps | 0 | 0 |
+
+MTP cut main forwards by about 52%, but increased expert payload by 40.7%; at
+this one-slot budget the extra draft/absorb/verify I/O made median decode 26.7%
+slower (displayed throughput 20.5% lower). Compressor occupancy fell by
+27-39 MB in the two complete MTP-off sidecar pairs and rose by only 46-185 MB
+in MTP-on cells. System swap was 0 before/after five complete cells; MTP-on
+repeat 3 ended at 0.75 MB with 48 system swapouts. Every process-level
+`time -l` counter remained zero, and there was no compression or swap spiral.
+
+### Authoritative repeated cold/warm matrix and exact TTFT
+
+The bounded harness speaks the length-prefixed serve protocol, sends one
+request, consumes END/STAT, sends RESET, repeats the exact prompt, then closes
+stdin. EOF exits the sole process; it is not a persistent server. Client TTFT
+is prompt-frame flush to first response byte; engine TTFT is BEGIN to the first
+token callback. They agree within 1.2 ms, far inside the enforced 50 ms bound.
+The final matrix is three independent processes per mode, two turns each:
+
+| Metric (median; range when variable) | MTP off, fresh-process turn | MTP off, LRU-warm turn | MTP on, fresh-process turn | MTP on, LRU-warm turn |
+|---|---:|---:|---:|---:|
+| Client TTFT | 29.332 s (29.154-29.730) | 28.431 s (28.058-28.498) | 31.132 s (31.095-31.234) | 31.895 s (31.640-31.936) |
+| Request elapsed | 190.848 s (186.445-196.556) | 189.761 s (181.318-190.210) | 246.331 s (245.942-247.490) | 241.362 s (225.342-245.478) |
+| STAT throughput | 0.34 tok/s (0.33-0.34) | 0.34 tok/s (0.34-0.35) | 0.26 tok/s | 0.27 tok/s (0.26-0.28) |
+| Hit rate | 1.9% | 2.0% | 1.1% | 1.1% |
+| Expert payload | 836.909 GB | 836.360 GB | 1126.703 GB | 1153.883 GB |
+| Main forwards / tok-forward | 63 / 1.02 | 63 / 1.02 | 30 / 2.13 | 31 / 2.06 |
+| MTP accepted / verified / raw | 0 / 0 / 0 | 0 / 0 / 0 | 34 / 87 / 90 | 33 / 90 / 93 |
+
+Median whole-process peak footprint was 13.631 GB MTP-off (range
+13.627-13.631) and 17.475 GB MTP-on (17.474-17.477). Every one of the six
+processes reported zero swaps. The same pre-existing 0.75 MB of system swap was
+present before and after every cell; none of the runs added to it.
+
+All twelve turns returned the exact same 64 token IDs and response SHA-256
+`494cf30efb6d4d592b5acfc4c8cc310756218bd14dddf3ea4342c9c521e89290`.
+MTP fresh-process acceptance lengths were `{0:11,1:9,2:5,3:5}` and warm were
+`{0:11,1:10,2:7,3:3}`. MTP reduced target forwards by about 52%, but its fresh
+turn took 29.1% longer and fetched 34.6% more expert payload than MTP-off; its
+warm turn took 27.2% longer and fetched 38.0% more. Preserving a one-slot LRU
+therefore does not fix the serialized expert-I/O tax, and MTP must be redesigned
+around overlap/residency before it can produce a net speedup in the port.
+
+The aggregate also pins the cache/I/O mechanism. Fresh MTP-off had median
+667.4 s cumulative read-service work, 117.5 s felt wait, 851 LRU hits, and
+44,245 load misses; fresh MTP-on rose to 969.5 s service, 159.1 s felt wait,
+664 hits, and 58,459 misses. Warm turns retained 75 experts (off) or 76 (on),
+1.4 GB total, with no pinned experts. These worker-cumulative service times
+can exceed request wall time because six loader workers overlap.
+
+An operator `mactop` snapshot during the MTP-on run independently showed the
+same bottleneck shape: about 4.9 GB/s disk reads and 0 B/s writes, only 24% GPU
+usage, 16.1 GB/s unified-memory bandwidth, 19.91/32 GB memory, and no swap.
+This is qualitative live telemetry rather than a matrix counter, but it strongly
+corroborates that current Colibri is feed/I/O limited on this M1 Max—not yet
+limited by GPU throughput, unified-memory bandwidth, or memory capacity.
+
+The deterministic aggregate is
+`runs/colibri-g0/results-20260722/warm-final-v4-matrix.json` (SHA-256
+`ca706bb22afd18554c7a80654c15b32c1578aeb06fbe847e9c8434bb6c8915af`);
+its rendered Markdown has SHA-256
+`fc8cc07613a9a5b97c120f7a2e4ddbf6111bbd99444a0ccae81e63e9d1caff90`.
+The JSON byte-counts and hashes every manifest, result, prompt, response,
+stdout, stderr, stats file, and pre/post system sidecar used by the matrix.
+
+### Real-model numeric oracle
+
+A measurement-only patch was applied only to the ignored exact-pin archive,
+never to `/Users/joshrossi/Code/colibri`. Its SHA-256 is
+`d9dfffe52c27d54cef0d89eee6a3cd865aed2bacf02e9d856cf31374903ac394`;
+the instrumented Metal binary is
+`3869b10708791d0c93b019f557d8917fa328b93047e977e132da6e316dcba84a`.
+Two runs emitted 140 records / about 54 MB covering sentinel GLM stages, MLA
+Q/compressed KV, sigmoid router scores and final top-8, main hidden/logits,
+MTP absorb/fusion/layer/head/logits, and the teacher token-16 decode. The two
+directories compare byte-for-byte. Main and MTP both choose token 16; their
+top1-top2 margins are 2.994461 and 3.650726. The tracked 163 KiB compact
+reduction is `fixtures/colibri-glm52/real-model-oracle.json`.
+
+### Artifact integrity and DSA decision
+
+Every one of 145 LFS payloads (383,760,044,154 bytes) matches the SHA-256 from
+exact revision `3cc8db99b1b13fc79325d987ba3c1c430766b3b8`; the evidence JSON
+SHA-256 is `518b03120daac07a42079edb29eb246ba636029a8b49091785d21519cc9ea939`.
+That schema-v2 evidence also records and validates the prior schema-v1 digest
+`c532e16f20f0fc4ba9d8e726215fd162c6fbb718d43388a06322304f4e006272`,
+so the metadata upgrade has an auditable chain rather than replacing the first
+full payload-hash result.
+Together with five non-LFS metadata files, all 150 artifact files are now
+accounted for.
+
+The public artifact contains no `out-idx-*` files or `indexer.*` tensors. A
+search found no stock-identical public overlay; the only structurally suitable
+published indexers derive from a different uncensored/fine-tuned model and are
+not valid exact-oracle evidence. Exact stock generation would selectively read
+20 source shards (99.90 GiB) from `zai-org/GLM-5.2-FP8` and needs a separate
+Torch environment. G0 therefore explicitly waives full-model DSA for this
+public-artifact baseline. The exact-pin model-free DSA fixture remains valid;
+a stock-generated overlay is deferred and must never mutate the HF snapshot.
+
+**G0 exit: met.** G1 may start. Raw ignored evidence is under
+`runs/colibri-g0/results-20260722/`; reproducible parsers/harnesses are
+`scripts/colibri-g0-{report,warm,oracle-report,matrix-report}.ts` and the
+artifact checker is `scripts/verify-colibri-g0-artifact.ts`.