akm-cli 0.9.0-beta.5 → 0.9.0-beta.51

package/dist/commands/improve/consolidate.js

@@ -5,52 +5,50 @@ import { createHash } from "node:crypto";
 import fs from "node:fs";
 import path from "node:path";
 import readline from "node:readline";
-import { parse as yamlParse } from "yaml";
+import consolidateSystemPrompt from "../../assets/prompts/consolidate-system.md" with { type: "text" };
 import { parseAssetRef } from "../../core/asset/asset-ref.js";
 import { assembleAssetFromString, serializeFrontmatter } from "../../core/asset/asset-serialize.js";
 import { parseFrontmatter } from "../../core/asset/frontmatter.js";
 import { resolveStashDir, timestampForFilename } from "../../core/common.js";
 import { getDefaultLlmConfig, loadConfig } from "../../core/config/config.js";
 import { ConfigError } from "../../core/errors.js";
-import { appendEvent } from "../../core/events.js";
+// Note: appendEvent import removed (WS-3a: archive TTL machinery retired)
 import { parseEmbeddedJsonResponse } from "../../core/parse.js";
+import { resolveStashStandards } from "../../core/standards/resolve-stash-standards.js";
 import { detectTruncatedDescription } from "../../core/text-truncation.js";
-import { hasHotCaptureMode, hasSupersededStatus, MERGE_ABSOLUTE_FLOOR_CHARS, MERGE_SHRINK_RATIO_MIN, validateProposalFrontmatter, } from "../proposal/validators/proposal-quality-validators.js";
+import { hasSupersededStatus, MERGE_ABSOLUTE_FLOOR_CHARS, MERGE_SHRINK_RATIO_MIN, validateProposalFrontmatter, } from "../proposal/validators/proposal-quality-validators.js";
 import { createProposal, isProposalSkipped, listProposals } from "../proposal/validators/proposals.js";
+import { cacheHash, runDeterministicDedup, stripFrontmatterBody } from "./dedup.js";
+import { checkGenerationGuard, checkLexicalDiversity, computeMergedGeneration, readAssetGeneration, runHomeostaticDemotion, shouldSkipHotProbationInLlm, } from "./homeostatic.js";
 import { writeContradictEdge } from "./memory/memory-belief.js";
 // Re-export the moved helpers so existing test imports continue to resolve.
 export { hasSupersededStatus, validateProposalFrontmatter };
+import { getBodyEmbeddings, getConsolidationJudgedMap, openStateDatabase, upsertBodyEmbeddings, upsertConsolidationJudged, withStateDb, } from "../../core/state-db.js";
 import { warn } from "../../core/warn.js";
 import { commitWriteTargetBoundary, deleteAssetFromSource, resolveWriteTarget, writeAssetToSource, } from "../../core/write-source.js";
 import { closeDatabase, findEntryIdByRef, getAllEntries, getEntryById, getNeighborsByEntryId, openExistingDatabase, } from "../../indexer/db/db.js";
 import { resolveImproveProcessRunnerFromProfile, runnerIsLlm } from "../../integrations/agent/runner.js";
 import { chatCompletion } from "../../llm/client.js";
-import { cosineSimilarity, embedBatch } from "../../llm/embedder.js";
+import { cosineSimilarity, embedBatch, resolveEmbeddingModelId } from "../../llm/embedder.js";
 import { isLlmFeatureEnabled, tryLlmFeature } from "../../llm/feature-gate.js";
+// Chunk sizing + per-chunk prompt assembly live in ./consolidate/chunking.
+// Imported for internal use by the orchestrator and re-exported for importers.
+import { buildChunkPrompt, computeSafeChunkSize, DEFAULT_CONTEXT_LENGTH_TOKENS } from "./consolidate/chunking.js";
+export { buildChunkPrompt, computeSafeChunkSize, DEFAULT_CONTEXT_LENGTH_TOKENS } from "./consolidate/chunking.js";
+// LLM-output sanitization (pure string/frontmatter transforms) lives in
+// ./consolidate/sanitize. Imported for internal use + re-exported for importers.
+import { normalizeUpdatedField, sanitizeMergedContent } from "./consolidate/sanitize.js";
+export { normalizeUpdatedField, sanitizeMergedContent, stripOuterCodeFence } from "./consolidate/sanitize.js";
+// Eligibility / safety predicates live in ./consolidate/eligibility. Imported
+// for internal guard use; the two public predicates are re-exported.
+import { consolidateGuardStatus, isConsolidationEligibleMemoryName, isHotCapturedMemory, isSessionCaptureMemoryName, } from "./consolidate/eligibility.js";
+export { isConsolidationEligibleMemoryName, isHotCapturedMemory, isSessionCaptureMemoryName, } from "./consolidate/eligibility.js";
+// Plan parsing / merging (pure op-reconciliation algebra) lives in
+// ./consolidate/merge. Imported for internal use; mergePlans re-exported.
+import { isValidOp, mergePlans } from "./consolidate/merge.js";
+export { mergePlans } from "./consolidate/merge.js";
 // ── Prompts ─────────────────────────────────────────────────────────────────
-const CONSOLIDATE_SYSTEM_PROMPT = `You are the akm consolidate assistant analyzing memory assets.
-
-Rules:
-1. MERGE: Two or more memories are substantially duplicated or closely related → propose merging. Return the primary ref to keep and secondary refs to delete. Do NOT include mergedContent — the merge will be executed in a separate step.
-2. DELETE: Memory is clearly outdated, contradicted, or redundant → propose deletion. NEVER propose delete for memories annotated \`(captureMode: hot)\` — they are user-explicit and only the user can retire them. The downstream guard will refuse these regardless, so proposing them just wastes tokens.
-3. PROMOTE: Memory expresses a stable, reusable fact suitable as a \`knowledge:\` asset → propose promotion. Do NOT delete the source memory. NEVER propose promote / merge / contradict for memories annotated \`(already queued)\` — they have a pending proposal whose body matches; a duplicate will be deterministically dropped, so proposing them just wastes tokens.
-4. CONTRADICT: Two memories make mutually exclusive factual claims about the same subject (e.g. "always use VPN" vs "VPN is optional") → mark the older or less authoritative one as contradicted. This writes a contradictedBy edge so the belief-resolution SCC algorithm can resolve the conflict. Do NOT delete contradicted memories — let the belief resolver decide.
-5. KEEP: Memory is unique and current → omit from output.
-
-Return ONLY JSON (no prose, no code fences):
-{
-  "operations": [
-    { "op": "merge", "primary": "memory:<name>", "secondaries": ["memory:<name>", ...], "mergeStrategy": "synthesize", "confidence": 0.95 },
-    { "op": "delete", "ref": "memory:<name>", "reason": "<brief reason>", "confidence": 0.90 },
-    { "op": "promote", "ref": "memory:<name>", "knowledgeRef": "knowledge:<suggested-slug>", "reason": "<brief reason>", "description": "<one sentence describing the new knowledge asset>", "confidence": 0.92 },
-    { "op": "contradict", "ref": "memory:<name>", "contradictedByRef": "memory:<name>", "reason": "<brief reason>", "confidence": 0.88 }
-  ],
-  "warnings": ["<optional concerns>"]
-}
-
-For every operation, emit a \`confidence\` field in [0, 1] expressing your certainty that the operation is correct and safe. Use 0.95+ only when evidence is unambiguous. Omit the field rather than guessing if you are uncertain.
-
-When the merged content includes an \`updated\` frontmatter field, the value MUST be a real ISO date string (e.g. \`updated: 2026-05-20\`). NEVER emit \`updated: today\`, \`updated: {today}\`, \`updated: {today: null}\`, \`updated: now\`, or any other literal placeholder/template-variable. If you do not have a real source-of-truth date, OMIT the \`updated\` field entirely — the post-processor will not invent one for you.`;
+const CONSOLIDATE_SYSTEM_PROMPT = consolidateSystemPrompt;
 /**
  * JSON Schema for structured consolidate plans (PR 1 of the asset-writers
  * decision — see knowledge:projects/akm/asset-writers-investigation/00-synthesis).
@@ -136,140 +134,17 @@ export const CONSOLIDATE_PLAN_JSON_SCHEMA = {
         },
     },
 };
-export function isConsolidationEligibleMemoryName(name) {
-    return !name.endsWith(".derived");
-}
-/**
- * Returns true when the memory file has `captureMode: hot` in its frontmatter.
- *
- * Hot memories are USER-EXPLICIT (written via `akm remember` on the hot path).
- * The consolidate LLM is forbidden from deleting or auto-merging them — the
- * user wrote them on purpose and only the user can decide to retire them.
- *
- * Reads the file once per check; consolidate runs against ~10 memories per
- * chunk so the IO cost is trivial. Returns false on any read/parse error
- * (fail-safe: an unparseable file is treated as not-hot, but the broader
- * consolidate flow already guards against unparseable memories elsewhere).
- *
- * Defends against four observed defect classes (see
- * `memory:akm-improve-critical-review-2026-05-20`):
- *   - LLM marks a memory contradicted then deletes (dangling contradictedBy)
- *   - LLM merges two unrelated memories sharing a topic keyword
- *   - LLM judges a recent durable design memo as "redundant"
- *   - Cascade deletes (LLM uses ref:X as `contradictedBy` for ref:Y then deletes both)
- */
-export function isHotCapturedMemory(filePath) {
-    try {
-        if (!fs.existsSync(filePath))
-            return false;
-        const content = fs.readFileSync(filePath, "utf8");
-        const parsed = parseFrontmatter(content);
-        return hasHotCaptureMode(parsed.data);
-    }
-    catch {
-        return false;
-    }
-}
-function consolidateGuardStatus(filePath) {
-    if (!fs.existsSync(filePath))
-        return "missing";
-    let content;
-    try {
-        content = fs.readFileSync(filePath, "utf8");
-    }
-    catch {
-        return "unparseable";
-    }
-    let parsed;
-    try {
-        parsed = parseFrontmatter(content);
-    }
-    catch {
-        return "unparseable";
-    }
-    const data = parsed.data;
-    if (!data || Object.keys(data).length === 0)
-        return "unparseable";
-    return hasHotCaptureMode(data) ? "hot" : "safe";
-}
-// ── Chunk sizing ─────────────────────────────────────────────────────────────
-/**
- * Conservative chars-per-token estimate used when computing prompt budgets.
- * English text averages roughly 4 chars/token for most LLM tokenizers. We use
- * 3 to stay conservative (shorter tokens = more tokens per char).
- */
-const CHARS_PER_TOKEN = 3;
-/**
- * Overhead budget reserved for the system prompt, chunk header lines, and per-
- * memory metadata lines (name, description, tags, separator). Measured at
- * roughly 600 chars for the system prompt + ~100 chars of header + ~50 chars
- * per memory × chunk size.  We round up to 2 000 tokens to leave room for the
- * model's own output.
- */
-const PROMPT_OVERHEAD_TOKENS = 2_000;
-/**
- * Default effective token budget used when the default LLM profile's
- * `contextLength` is not set. This is intentionally conservative (4 096)
- * rather than being set to the model's actual context window, because:
- *
- *   - When the agent path is used, the agent CLI (e.g. opencode)
- *     prepends its own large system prompt + conversation history before
- *     forwarding to the model. That overhead easily consumes 30K+ tokens on
- *     a model with a 16K context window, leaving very little room for
- *     chunk content.
- *   - When the HTTP path is used (an LLM profile is selected), only the akm
- *     system prompt and user prompt are sent, so the budget can be set to the
- *     model's actual context length via profiles.llm[defaults.llm].contextLength.
- *
- * Set profiles.llm[defaults.llm].contextLength in your config file to the
- * model's actual context window to allow larger chunks on the HTTP path.
- */
-export const DEFAULT_CONTEXT_LENGTH_TOKENS = 4_096;
-/**
- * Given the model's context window and the per-memory body truncation limit,
- * return the maximum number of memories that can safely fit in one chunk
- * without the prompt overflowing the context window.
- *
- * The formula is:
- *   usableTokens = contextLength - PROMPT_OVERHEAD_TOKENS
- *   tokensPerMemory = ceil(bodyTruncation / CHARS_PER_TOKEN)
- *   chunkSize = floor(usableTokens / tokensPerMemory)
- *
- * Result is clamped between 1 and 50 to avoid degenerate values.
- *
- * @param contextLength - Model context window in tokens.
- * @param bodyTruncation - Max chars per memory body included in the prompt.
- * @param maxChunkSize - Optional override for the hardcoded cap of 50 (1–50).
- */
-export function computeSafeChunkSize(contextLength, bodyTruncation, maxChunkSize) {
-    const usableTokens = Math.max(contextLength - PROMPT_OVERHEAD_TOKENS, 0);
-    const tokensPerMemory = Math.max(Math.ceil(bodyTruncation / CHARS_PER_TOKEN), 1);
-    const raw = Math.floor(usableTokens / tokensPerMemory);
-    return Math.max(1, Math.min(maxChunkSize ?? 50, raw));
-}
-// ── Similarity clustering (C-1 / #380) ──────────────────────────────────────
-/**
- * Re-order memories so that similar ones are placed adjacent to each other
- * before the memories are sliced into chunks. This ensures high-similarity
- * memories land in the same LLM context window, allowing the consolidate
- * model to detect and merge duplicates that would otherwise be split across
- * chunks and survive indefinitely.
- *
- * Algorithm: greedy nearest-neighbour chain starting from the first memory.
- * Each step selects the unused memory with the highest cosine similarity to
- * the last-placed memory. O(n²) — acceptable for the expected N < 200.
- *
- * mem0 arXiv:2504.19413 — every candidate compared against whole store.
- * A-MEM arXiv:2502.12110 — atomic notes linked by similarity.
- *
- * Returns the original order unchanged when:
- *   - The embedding config is not present.
- *   - Embedding requests fail (fail-open).
- *   - There are fewer than 3 memories (no benefit to reordering).
- */
-async function clusterMemoriesBySimilarity(memories, config) {
+async function clusterMemoriesBySimilarity(memories, config, stateDb) {
+    const noTelemetry = { embedMs: 0, cacheHits: 0, cacheMisses: 0 };
     if (memories.length < 3 || !config.embedding)
-        return memories;
+        return { ordered: memories, embedTelemetry: noTelemetry };
+    // WS-3a: cluster uses description+tags as the embedding input (NOT the raw
+    // body) — this is intentionally different from the dedup/body cache because
+    // the clustering goal is semantic grouping, not dedup twin detection.
+    // The body_embeddings cache is keyed by cacheHash(body); clustering inputs
+    // are keyed by cacheHash(description+tags text). Re-use the same table with
+    // a distinct hash so the two lookup sets never collide.
+    const modelId = resolveEmbeddingModelId(config.embedding);
     const texts = memories.map((m) => {
         const parts = [];
         if (m.description)
@@ -278,16 +153,91 @@ async function clusterMemoriesBySimilarity(memories, config) {
             parts.push(m.tags.join(" "));
         return parts.join(". ") || m.name;
     });
-    let embeddings = null;
-    try {
-        embeddings = await embedBatch(texts, config.embedding);
+    // Compute content hashes for the cluster texts (not bodies — different input).
+    const contentHashes = texts.map((t) => createHash("sha256").update(t, "utf8").digest("hex"));
+    // WS-5: track embed cache hits/misses for perf telemetry.
+    let embedMs = 0;
+    let cacheHits = 0;
+    let cacheMisses = 0;
+    let cachedVecs = new Map();
+    if (stateDb) {
+        try {
+            cachedVecs = getBodyEmbeddings(stateDb, contentHashes, modelId);
+        }
+        catch {
+            // Fail open.
+            cachedVecs = new Map();
+        }
     }
-    catch {
-        // Fail open: embedding failures degrade gracefully to original order.
-        return memories;
+    const missIndices = [];
+    const missTexts = [];
+    for (let i = 0; i < texts.length; i++) {
+        if (!cachedVecs.has(contentHashes[i])) {
+            missIndices.push(i);
+            missTexts.push(texts[i]);
+            cacheMisses++;
+        }
+        else {
+            cacheHits++;
+        }
     }
+    let missVecs = [];
+    if (missTexts.length > 0) {
+        const embedStart = Date.now();
+        try {
+            missVecs = await embedBatch(missTexts, config.embedding);
+        }
+        catch {
+            // Fail open: embedding failures degrade gracefully to original order.
+            return { ordered: memories, embedTelemetry: { embedMs, cacheHits, cacheMisses } };
+        }
+        finally {
+            embedMs += Date.now() - embedStart;
+        }
+        // Upsert newly computed vectors into the cache.
+        if (stateDb && missVecs.length === missTexts.length) {
+            try {
+                const toUpsert = missIndices.map((idx, pos) => ({
+                    contentHash: contentHashes[idx],
+                    embedding: missVecs[pos],
+                    modelId,
+                }));
+                upsertBodyEmbeddings(stateDb, toUpsert);
+            }
+            catch {
+                // Fail open: cache write errors are non-fatal.
+            }
+        }
+    }
+    // Assemble the full embedding array in memories order.
+    let embeddings = null;
+    {
+        const assembled = [];
+        let ok = true;
+        for (let i = 0; i < memories.length; i++) {
+            const hash = contentHashes[i];
+            const cached = cachedVecs.get(hash);
+            if (cached) {
+                assembled.push(cached);
+                continue;
+            }
+            const missPos = missIndices.indexOf(i);
+            const vec = missPos >= 0 ? missVecs[missPos] : undefined;
+            if (vec) {
+                assembled.push(vec);
+            }
+            else {
+                ok = false;
+                break;
+            }
+        }
+        if (ok && assembled.length === memories.length) {
+            embeddings = assembled;
+        }
+    }
+    const embedTelemetry = { embedMs, cacheHits, cacheMisses };
     if (!embeddings || embeddings.length !== memories.length)
-        return memories;
+        return { ordered: memories, embedTelemetry };
     // Greedy nearest-neighbour chain.
     const used = new Array(memories.length).fill(false);
     const ordered = [];
@@ -313,87 +263,16 @@ async function clusterMemoriesBySimilarity(memories, config) {
         used[bestIdx] = true;
         current = bestIdx;
     }
-    return ordered;
+    return { ordered, embedTelemetry };
 }
 // ── Chunk helpers ────────────────────────────────────────────────────────────
-/**
- * Build the per-chunk user prompt fed to the consolidate LLM.
- *
- * Each memory is annotated with two flags that drive the system-prompt
- * rules at lines 181-186:
- *   - `(captureMode: hot)` — user-explicit memory; system prompt rule 2
- *     forbids proposing delete. ~60 wasted LLM verdicts/4h on this user's
- *     stack before this annotation.
- *   - `(already queued)` — the memory's body hash matches a pending
- *     consolidate proposal; system prompt rule 3 forbids proposing
- *     promote/merge/contradict. ~107/4h before this annotation.
- *
- * Both annotations are visible to the LLM. `pendingProposalBodyHashes`
- * is precomputed once per run by `loadPendingConsolidateProposalHashes`
- * so the cost stays O(memories) inside the chunk loop.
- */
-export function buildChunkPrompt(sourceName, memories, chunkIndex, totalChunks, bodyTruncation, pendingProposalBodyHashes = new Set()) {
-    const start = memories[0] ? `memory:${memories[0].name}` : "";
-    const end = memories[memories.length - 1] ? `memory:${memories[memories.length - 1].name}` : "";
-    const annotationsByIndex = [];
-    const hotRefs = [];
-    for (const m of memories) {
-        let body = "";
-        try {
-            body = fs.readFileSync(m.filePath, "utf8");
-        }
-        catch {
-            body = "(unreadable)";
-        }
-        const parsed = parseFrontmatter(body);
-        const isHot = parsed.data.captureMode === "hot";
-        const bodyHash = createHash("sha256").update(parsed.content.trim(), "utf8").digest("hex");
-        const isAlreadyQueued = pendingProposalBodyHashes.has(bodyHash);
-        annotationsByIndex.push({ isHot, isAlreadyQueued, body });
-        if (isHot)
-            hotRefs.push(`memory:${m.name}`);
-    }
-    const lines = [
-        `Source: ${sourceName}`,
-        `Chunk ${chunkIndex + 1} of ${totalChunks}, memories ${start}–${end}:`,
-        "",
-    ];
-    // Top-of-prompt protection block for hot refs. Neutral phrasing — avoid
-    // op-words like "promote", "merge", "contradict" so the model doesn't
-    // accidentally treat the warning as a hint to use that op elsewhere
-    // (variant B leaked the word "contradict" into the control sample
-    // during the diagnostic).
-    if (hotRefs.length > 0) {
-        lines.push("⛔ DO NOT propose any `delete` operation for these refs — they are user-explicit (captureMode: hot) and the downstream guard refuses them regardless. Proposing delete for any of these only wastes tokens.");
-        for (const ref of hotRefs)
-            lines.push(`  - ${ref}`);
-        lines.push("");
-    }
-    for (let i = 0; i < memories.length; i++) {
-        const m = memories[i];
-        const { isHot, isAlreadyQueued, body } = annotationsByIndex[i];
-        const annotations = [];
-        if (isHot)
-            annotations.push("captureMode: hot");
-        if (isAlreadyQueued)
-            annotations.push("already queued");
-        const annotationSuffix = annotations.length > 0 ? ` (${annotations.join("; ")})` : "";
-        lines.push(`[${i + 1}] memory:${m.name}${annotationSuffix}`);
-        lines.push(`Description: ${m.description || "(none)"}`);
-        lines.push(`Tags: ${m.tags.length > 0 ? m.tags.join(", ") : "(none)"}`);
-        lines.push("---");
-        lines.push(body.slice(0, bodyTruncation));
-        lines.push("");
-    }
-    return lines.join("\n");
-}
 /**
  * Precompute body-hashes of all currently-pending consolidate proposals so
  * the per-chunk prompt can annotate memories whose body would just produce
- * a deterministic `dedup_pending_proposal` skip. Hash domain matches the
- * dedup site at ~line 1510 (sha256 over the post-frontmatter content,
- * trimmed). Empty set on any read/parse error — fail-safe to "annotate
- * nothing" so the LLM still proposes, just slightly more wastefully.
+ * a deterministic `dedup_pending_proposal` skip. Uses `cacheHash` (case-
+ * preserving stripped body) — the same domain used by the body-embedding
+ * cache and `computeMemoryContentHash`. Empty set on any read/parse error
+ * — fail-safe to "annotate nothing" so the LLM still proposes.
  */
 function loadPendingConsolidateProposalHashes(stashDir) {
     const hashes = new Set();
@@ -401,8 +280,7 @@ function loadPendingConsolidateProposalHashes(stashDir) {
         const pending = listProposals(stashDir, { status: "pending" }).filter((p) => p.source === "consolidate");
         for (const p of pending) {
             try {
-                const body = parseFrontmatter(p.payload.content).content.trim();
-                hashes.add(createHash("sha256").update(body, "utf8").digest("hex"));
+                hashes.add(cacheHash(p.payload.content));
             }
             catch {
                 // skip malformed payloads — they can't dedup anyway
@@ -414,148 +292,6 @@ function loadPendingConsolidateProposalHashes(stashDir) {
     }
     return hashes;
 }
-function isValidOp(op) {
-    if (typeof op !== "object" || op === null)
-        return false;
-    const o = op;
-    if (o.op === "merge") {
-        return typeof o.primary === "string" && Array.isArray(o.secondaries);
-    }
-    if (o.op === "delete") {
-        return typeof o.ref === "string";
-    }
-    if (o.op === "promote") {
-        return typeof o.ref === "string" && typeof o.knowledgeRef === "string";
-    }
-    if (o.op === "contradict") {
-        return typeof o.ref === "string" && typeof o.contradictedByRef === "string";
-    }
-    return false;
-}
-export function mergePlans(chunks, knownRefs) {
-    const mergeOps = new Map();
-    const deleteOps = new Map();
-    const promoteOps = new Map();
-    // C-3 / #382: contradict ops keyed by `ref|contradictedByRef` to deduplicate.
-    const contradictOps = new Map();
-    const warnings = [];
-    for (const chunk of chunks) {
-        for (const op of chunk) {
-            if (op.op === "merge") {
-                // Drop ops whose primary the LLM hallucinated (not in the loaded memory
-                // pool). Without this guard, a hallucinated primary flows all the way to
-                // Phase B where !memoryByRef.has(primary) fires and charges every real
-                // secondary with merge_primary_missing — masking LLM hallucinations as
-                // filter regressions in health metrics.
-                if (knownRefs && !knownRefs.has(op.primary)) {
-                    warnings.push(`mergePlans: primary ${op.primary} not in loaded memory pool (LLM hallucination) — dropping op before execution.`);
-                    // Use a dedicated skip reason so dashboards can distinguish
-                    // hallucinated primaries from stale-DB regressions.
-                    // Secondaries are real refs; they are NOT charged here — they remain
-                    // available for other ops to claim.
-                    continue;
-                }
-                // Filter hallucinated secondaries while preserving real ones.
-                let mergeOp = op;
-                if (knownRefs) {
-                    const filteredSecondaries = op.secondaries.filter((sec) => {
-                        if (!knownRefs.has(sec)) {
-                            warnings.push(`mergePlans: secondary ${sec} not in loaded memory pool (LLM hallucination) — dropping from op.`);
-                            return false;
-                        }
-                        return true;
-                    });
-                    if (filteredSecondaries.length !== op.secondaries.length) {
-                        mergeOp = { ...op, secondaries: filteredSecondaries };
-                    }
-                }
-                // merge wins over delete
-                if (deleteOps.has(mergeOp.primary)) {
-                    deleteOps.delete(mergeOp.primary);
-                }
-                for (const sec of mergeOp.secondaries) {
-                    if (deleteOps.has(sec))
-                        deleteOps.delete(sec);
-                }
-                mergeOps.set(mergeOp.primary, mergeOp);
-            }
-            else if (op.op === "delete") {
-                // merge and promote both win over delete. A promote is non-destructive
-                // (creates a proposal) but the source memory is counted in `promoted`;
-                // if a delete also fires, the ref lands in both `promoted` and
-                // `skipReasons`, breaking the invariant by +1.
-                if (!mergeOps.has(op.ref) && !promoteOps.has(op.ref)) {
-                    deleteOps.set(op.ref, op);
-                }
-            }
-            else if (op.op === "promote") {
-                // C-2 / #381: when both a promote and a merge target the same ref,
-                // queue the promote FIRST rather than discarding it. The promote op
-                // routes through createProposal (the human-gated proposal queue), so
-                // it is non-destructive. The merge follows after the proposal is
-                // created. This preserves the human reviewer's ability to inspect the
-                // promotion before the source memory is merged/deleted.
-                // AGM K*8 — retain the maximally informative consistent subset.
-                promoteOps.set(op.ref, op);
-            }
-            else if (op.op === "contradict") {
-                // Deduplicate by ref+contradictedByRef pair.
-                const key = `${op.ref}|${op.contradictedByRef}`;
-                if (!contradictOps.has(key)) {
-                    contradictOps.set(key, op);
-                }
-            }
-        }
-    }
-    // Second pass: enforce merge-wins-over-delete and deduplicate secondaries.
-    //
-    // 1. Delete/secondary ordering bug: the per-chunk loop removes delete ops
-    //    for secondaries that were already in deleteOps, but misses the case
-    //    where the delete chunk came first. A full sweep here fixes both orders.
-    //
-    // 2. Cross-merge secondary dedup: if ref A is a secondary in two merge ops,
-    //    only the first (insertion-order) retains it. Without this, a successful
-    //    merge credits A to mergedSecondaries and a later merge's emitMerge-
-    //    FailureSkips also charges A to skipReasons — double-counting A while
-    //    processed has it only once.
-    //
-    // 3. Primary-as-secondary dedup: if ref A is a primary in one merge op and
-    //    a secondary in another, remove A from the secondary list. Both merges
-    //    would otherwise claim A (merged++ for A, then mergedSecondaries++ for A)
-    //    breaking the invariant the same way.
-    // Also remove delete ops for any ref claimed by a promote op (handles the
-    // case where the delete chunk appeared before the promote chunk).
-    for (const ref of promoteOps.keys()) {
-        deleteOps.delete(ref);
-    }
-    const claimedSecondaries = new Set();
-    for (const mergeOp of mergeOps.values()) {
-        deleteOps.delete(mergeOp.primary);
-        mergeOp.secondaries = mergeOp.secondaries.filter((sec) => {
-            if (mergeOps.has(sec)) {
-                warnings.push(`Merge: secondary ${sec} is also a merge primary — removing from secondary list to avoid double-count.`);
-                return false;
-            }
-            if (claimedSecondaries.has(sec)) {
-                warnings.push(`Merge: secondary ${sec} appears in multiple merge ops — retaining in first op only.`);
-                return false;
-            }
-            claimedSecondaries.add(sec);
-            deleteOps.delete(sec);
-            return true;
-        });
-    }
-    // C-2 / #381: promote ops are ordered BEFORE merge ops so that the
-    // human-gated proposal queue entry is created before any destructive merge.
-    // Phase B processes ops in array order, so promote executes first.
-    const ops = [
-        ...promoteOps.values(),
-        ...mergeOps.values(),
-        ...deleteOps.values(),
-        ...contradictOps.values(),
-    ];
-    return { ops, warnings };
-}
 function getJournalPath(stashDir) {
     return path.join(stashDir, ".akm", "consolidate-journal.json");
 }
@@ -672,6 +408,28 @@ function backupFile(filePath, backupDir, name) {
         // best-effort
     }
 }
+// ── WS-3b: Generation frontmatter injection ───────────────────────────────────
+/**
+ * Inject `generation` (and optionally `source_refs`) into merged content.
+ * generation = max(sourceGenerations) + 1.
+ * Fails open — returns original content if frontmatter can't be parsed.
+ */
+function injectGenerationFrontmatter(mergedContent, sourceGenerations, allParticipants) {
+    try {
+        const parsed = parseFrontmatter(mergedContent);
+        const updatedFm = {
+            ...parsed.data,
+            generation: computeMergedGeneration(sourceGenerations),
+        };
+        if (!updatedFm.source_refs) {
+            updatedFm.source_refs = allParticipants;
+        }
+        return assembleAssetFromString(serializeFrontmatter(updatedFm), parsed.content);
+    }
+    catch {
+        return mergedContent; // fail open
+    }
+}
 // ── Archive helper (P1-B: soft-invalidation) ─────────────────────────────────
 /**
  * Move a memory asset to `.akm/archive/` with `status: superseded` frontmatter
@@ -752,6 +510,25 @@ function resolveConsolidateLlmConfig(config) {
     // fall back to the default LLM profile rather than disabling the pass.
     return getDefaultLlmConfig(config);
 }
+// ── Judged-state cache (#581) ────────────────────────────────────────────────
+/**
+ * Stable content hash for a memory file used by the judged-state cache (#581)
+ * and the body-embedding cache (WS-3a). Uses `cacheHash` from dedup.ts:
+ * sha256 of the case-preserving stripped body. Two memories that differ only
+ * in frontmatter (`updated:`, `inferenceProcessed:`) hash identically, so a
+ * cosmetic frontmatter touch never forces a needless re-judge — only a body
+ * change does. Returns `undefined` on any read/parse error so callers fail
+ * open (treat the memory as un-cached → it stays in the LLM pool).
+ */
+function computeMemoryContentHash(filePath) {
+    try {
+        const raw = fs.readFileSync(filePath, "utf8");
+        return cacheHash(raw);
+    }
+    catch {
+        return undefined;
+    }
+}
 // ── Main entry point ─────────────────────────────────────────────────────────
 export async function akmConsolidate(opts = {}) {
     const startMs = Date.now();
@@ -780,6 +557,27 @@ export async function akmConsolidate(opts = {}) {
     }
     const warnings = [];
     checkForIncompleteJournal(stashDir, opts.recoveryMode ?? "abort", warnings);
+    // WS-3a: open one state.db handle shared by the body-embedding cache (dedup
+    // + cluster) and the judged-state cache. All callers in the function body
+    // receive this handle; it is closed in the `finally` block below.
+    // Fail-open: any open error leaves it `undefined` and all cache paths skip.
+    let sharedStateDb;
+    try {
+        sharedStateDb = openStateDatabase();
+    }
+    catch {
+        // State DB unavailable → skip the embedding cache for this run.
+    }
+    try {
+        return await akmConsolidateInner(opts, config, stashDir, startMs, sourceRun, warnings, sharedStateDb);
+    }
+    finally {
+        sharedStateDb?.close();
+    }
+}
+// Inner implementation — all early-return paths are here; sharedStateDb is
+// closed by the outer finally in `akmConsolidate`.
+async function akmConsolidateInner(opts, config, stashDir, startMs, sourceRun, warnings, sharedStateDb) {
     let memories = loadMemoriesForSource(opts.target, stashDir, warnings);
     // Pre-flight: filter out stale DB entries whose files no longer exist on
     // disk. Without this, memories deleted by a prior run (but not yet
@@ -791,6 +589,80 @@ export async function akmConsolidate(opts = {}) {
         warnings.push(`Pre-flight: filtered ${staleCount} stale DB entr${staleCount === 1 ? "y" : "ies"} (file absent on disk) from memory pool before chunking.`);
     }
     memories = memories.filter((m) => fs.existsSync(m.filePath));
+    // ── WS-3b Step 0a: Homeostatic demotion ────────────────────────────────────
+    // DEFAULT OFF. Before any LLM merge, demote retrievalSalience in state.db
+    // for stale/low-value assets so the merge pool is bounded and high-SNR.
+    // Demotion is state.db-only (file content untouched); re-promotable on
+    // re-retrieval. Only fires when `homeostaticDemotion.enabled === true`.
+    const homeostaticConfig = config.profiles?.improve?.default?.processes?.consolidate?.homeostaticDemotion ?? {};
+    if (homeostaticConfig.enabled && sharedStateDb) {
+        const demotionResult = runHomeostaticDemotion(sharedStateDb, homeostaticConfig);
+        if (demotionResult.demoted > 0) {
+            warnings.push(`Homeostatic demotion: demoted retrievalSalience for ${demotionResult.demoted} stale asset(s) before merge pool assembly.`);
+        }
+        warnings.push(...demotionResult.warnings);
+    }
+    // ── WS-3b Step 0c: Filter hot-probation assets from LLM merge pool ─────────
+    // Hot-probation assets (system-generated, not yet graduated from intake pass)
+    // are processed by the dedup pre-pass but excluded from the LLM clustering.
+    // This prevents noisy extractions from polluting LLM context. The dedup pass
+    // below still runs against them so they're cleaned up deterministically.
+    // DEFAULT OFF — only active when `processes.extract.hotProbation.enabled === true`
+    // (the flag that causes extract to tag new extractions as hot-probation).
+    // Without that flag no assets will ever carry the hot-probation marker, so
+    // running the filter loop would be pure unnecessary I/O over the full corpus.
+    const hotProbationEnabled = config.profiles?.improve?.default?.processes?.extract?.hotProbation
+        ?.enabled === true;
+    let hotProbationCount = 0;
+    if (hotProbationEnabled) {
+        const hotProbationMemories = [];
+        const nonProbationMemories = [];
+        for (const m of memories) {
+            try {
+                const raw = fs.readFileSync(m.filePath, "utf8");
+                const parsed = parseFrontmatter(raw);
+                if (shouldSkipHotProbationInLlm(parsed.data)) {
+                    hotProbationMemories.push(m);
+                    hotProbationCount++;
+                }
+                else {
+                    nonProbationMemories.push(m);
+                }
+            }
+            catch {
+                nonProbationMemories.push(m); // fail open
+            }
+        }
+        if (hotProbationCount > 0) {
+            warnings.push(`Hot-probation: ${hotProbationCount} hot-probation asset(s) routed to dedup-only pass (excluded from LLM merge pool).`);
+            memories = nonProbationMemories;
+        }
+    }
+    // ── Deterministic dedup pre-pass (#617) ─────────────────────────────────────
+    // Cheap, no-LLM fast path that collapses the obvious near-duplicates
+    // (`.derived` ↔ origin pairs + content twins) BEFORE the embedding-clustered
+    // LLM consolidation. DEFAULT OFF — when `dedup.enabled !== true` this is a
+    // no-op and the pass behaves byte-identically to today. Collapsed variants
+    // are pruned from the LLM pool so the model only ever sees genuinely
+    // distinct-but-related memories. Each dropped variant is archived (soft
+    // invalidation) before deletion, matching the LLM merge path.
+    // Dry-run never mutates the filesystem, so the dedup pre-pass is skipped
+    // entirely under `--dry-run` (the LLM plan preview below is unaffected).
+    let dedupCollapsed = 0;
+    if (opts.dedup?.enabled && !opts.dryRun) {
+        const dedupTimestamp = timestampForFilename();
+        const dedupResult = await runDeterministicDedup(stashDir, opts.dedup, config, (variantFilePath, variantName) => {
+            archiveMemory(variantFilePath, stashDir, `memory:${variantName}`, "collapsed by deterministic dedup pre-pass", -1, undefined, warnings);
+            backupFile(variantFilePath, getBackupDir(stashDir, dedupTimestamp), variantName);
+        }, opts.signal, sharedStateDb);
+        dedupCollapsed = dedupResult.collapsed;
+        warnings.push(...dedupResult.warnings);
+        if (dedupResult.consumedRefs.length > 0) {
+            const consumed = new Set(dedupResult.consumedRefs);
+            memories = memories.filter((m) => !consumed.has(`memory:${m.name}`));
+            warnings.push(`Deterministic dedup: collapsed ${dedupResult.collapsed} near-duplicate memor${dedupResult.collapsed === 1 ? "y" : "ies"} (no LLM) before chunking.`);
+        }
+    }
     if (memories.length === 0) {
         return {
             schemaVersion: 1,
@@ -801,7 +673,10 @@ export async function akmConsolidate(opts = {}) {
             target: opts.target ?? stashDir,
             processed: 0,
             merged: 0,
-            deleted: 0,
+            // #617: the deterministic dedup pre-pass may have emptied the pool by
+            // collapsing every remaining memory into a canonical. Surface those
+            // collapses in `deleted` so the run reports the work it actually did.
+            deleted: dedupCollapsed,
             promoted: [],
             contradicted: 0,
             warnings,
@@ -809,7 +684,7 @@ export async function akmConsolidate(opts = {}) {
         };
     }
     if (opts.incrementalSince) {
-        memories = narrowToIncrementalCandidates(memories, opts.incrementalSince, warnings);
+        memories = narrowToIncrementalCandidates(memories, opts.incrementalSince, warnings, opts.neighborsPerChanged);
         if (memories.length === 0) {
             return {
                 schemaVersion: 1,
@@ -828,6 +703,109 @@ export async function akmConsolidate(opts = {}) {
             };
         }
     }
+    // WS-5 perf telemetry accumulators. These are collected throughout the run and
+    // merged into `perfTelemetry` on the final ConsolidateResult.
+    // `dedupPoolSize` = memories entering judgedCache narrowing (after dedup+incremental+limit).
+    // `judgedCacheSkipped` = memories skipped by the cache.
+    // `llmPoolSize` = memories actually sent to the LLM.
+    // `embedMs/cacheHits/cacheMisses` = accumulated from clusterMemoriesBySimilarity.
+    const perfMs = { dedupPoolSize: memories.length, judgedCacheSkipped: 0 };
+    // ── Judged-state cache narrowing (#581) ─────────────────────────────────────
+    // DEFAULT OFF. When enabled, skip every memory whose current content hash
+    // equals the hash recorded the last time the consolidate LLM judged it
+    // (judged-unchanged → no re-judge). This converts coverage from O(window) to
+    // O(changed/new) so one run can sweep the whole corpus while the LLM only
+    // sees genuinely new/changed memories. `currentHashByName` is populated for
+    // EVERY surviving memory (whether or not the cache is on) so the post-LLM
+    // recording step can upsert judged state without re-reading the files; when
+    // the cache is off it stays empty and the recording step is a no-op.
+    const judgedCacheEnabled = opts.judgedCache?.enabled !== false;
+    const currentHashByName = new Map();
+    if (judgedCacheEnabled) {
+        for (const m of memories) {
+            const h = computeMemoryContentHash(m.filePath);
+            if (h !== undefined)
+                currentHashByName.set(m.name, h);
+        }
+        let cachedMap = new Map();
+        {
+            // Use the shared state.db handle if available; open a local one otherwise.
+            const dbForJudged = sharedStateDb;
+            if (dbForJudged) {
+                try {
+                    cachedMap = getConsolidationJudgedMap(dbForJudged, memories.map((m) => `memory:${m.name}`));
+                }
+                catch {
+                    cachedMap = new Map();
+                }
+            }
+            else {
+                try {
+                    cachedMap = withStateDb((localDb) => getConsolidationJudgedMap(localDb, memories.map((m) => `memory:${m.name}`)));
+                }
+                catch {
+                    // State DB unavailable → fail open: judge the full pool this run.
+                    cachedMap = new Map();
+                }
+            }
+        }
+        const beforeCount = memories.length;
+        memories = memories.filter((m) => {
+            const cur = currentHashByName.get(m.name);
+            // No readable hash → keep (fail open; let the LLM judge it).
+            if (cur === undefined)
+                return true;
+            const cached = cachedMap.get(`memory:${m.name}`);
+            // Skip only when previously judged AND content is byte-identical since.
+            return !(cached !== undefined && cached.content_hash === cur);
+        });
+        const skipped = beforeCount - memories.length;
+        perfMs.judgedCacheSkipped = skipped; // WS-5 perf telemetry
+        if (skipped > 0) {
+            warnings.push(`Judged-state cache: skipped ${skipped} memor${skipped === 1 ? "y" : "ies"} judged-unchanged (no LLM); ${memories.length} remain for judging.`);
+        }
+        if (memories.length === 0) {
+            return {
+                schemaVersion: 1,
+                ok: true,
+                shape: "consolidate-result",
+                dryRun: opts.dryRun ?? false,
+                previewOnly: false,
+                target: opts.target ?? stashDir,
+                processed: 0,
+                merged: 0,
+                deleted: dedupCollapsed,
+                promoted: [],
+                contradicted: 0,
+                warnings,
+                durationMs: Date.now() - startMs,
+            };
+        }
+    }
+    if (opts.limit === undefined && memories.length > 150) {
+        warnings.push(`Consolidation: pool has ${memories.length} memories and no limit is set. Consider adding a limit to your consolidate config to prevent timeouts on slow LLM endpoints.`);
+    }
+    if (opts.limit !== undefined && memories.length > opts.limit) {
+        // Order oldest-modified-first before capping so the limit selects the
+        // stalest memories rather than a fixed head of the (rowid-ordered) DB
+        // query. Consolidation rewrites surviving files, bumping their mtime, so
+        // processed memories drift to the back of the queue and the cap rotates
+        // across the whole corpus over successive runs instead of revisiting the
+        // same slice every time. Fail-open to 0 (front of queue) when a file can
+        // no longer be stat'd.
+        const mtimeOf = (m) => {
+            try {
+                return fs.statSync(m.filePath).mtimeMs;
+            }
+            catch {
+                return 0;
+            }
+        };
+        const mtimeCache = new Map(memories.map((m) => [m.filePath, mtimeOf(m)]));
+        memories = [...memories].sort((a, b) => (mtimeCache.get(a.filePath) ?? 0) - (mtimeCache.get(b.filePath) ?? 0));
+        warnings.push(`Consolidation: pool capped at ${opts.limit} of ${memories.length} memories (limit option, oldest-modified first).`);
+        memories = memories.slice(0, opts.limit);
+    }
     // Consolidation always uses the HTTP LLM client directly — never the agent
     // CLI. The agent CLI is for interactive agent sessions (reflect, propose);
     // structured JSON generation works better and faster via HTTP.
@@ -851,16 +829,64 @@ export async function akmConsolidate(opts = {}) {
     const chunkSize = computeSafeChunkSize(modelContextLength, bodyTruncation, opts.maxChunkSize);
     // -- Phase A: plan generation -----------------------------------------------
     const sourceName = opts.target ?? stashDir;
+    // WS-5: capture llmPoolSize = memories entering the LLM (after all filtering).
+    const llmPoolSize = memories.length;
     // C-1 / #380: Pre-cluster memories by embedding similarity before chunking.
     // This ensures that semantically similar memories land in the same LLM
     // context window, allowing the model to detect and merge duplicates that
     // would otherwise be split across chunks and survive indefinitely.
     // mem0 arXiv:2504.19413, A-MEM arXiv:2502.12110.
     // Fails open: if embeddings are unavailable or fail, original order is used.
-    const clusteredMemories = await clusterMemoriesBySimilarity(memories, config);
+    const { ordered: clusteredMemories, embedTelemetry } = await clusterMemoriesBySimilarity(memories, config, sharedStateDb);
+    // WS-3b Anti-collapse step 8c: inject random (non-similar) clusters.
+    // A small fraction (default 5%) of the pool is shuffled into random positions
+    // so the pipeline isn't PURELY similarity-driven. This prevents rich-get-richer
+    // entrenchment where only the most-retrieved assets ever get consolidated.
+    // DEFAULT OFF — gated on antiCollapse.enabled.
+    let finalClusteredMemories = clusteredMemories;
+    {
+        const antiCollapseForCluster = config.profiles?.improve?.default?.processes?.consolidate?.antiCollapse ?? {};
+        if (antiCollapseForCluster.enabled && clusteredMemories.length > 2) {
+            const fraction = antiCollapseForCluster.randomClusterFraction ?? 0.05;
+            const randomCount = Math.max(1, Math.floor(clusteredMemories.length * fraction));
+            // Pick `randomCount` positions to inject random (un-clustered) members.
+            // Use a seeded-ish shuffle: sort by hash of the name so it's deterministic
+            // per run but not strictly similarity-driven.
+            const shuffled = [...clusteredMemories].sort((a, b) => {
+                // Deterministic shuffle: compare sha256-ish (use name hash as proxy).
+                const ha = a.name.split("").reduce((acc, c) => ((acc << 5) - acc + c.charCodeAt(0)) | 0, 0);
+                const hb = b.name.split("").reduce((acc, c) => ((acc << 5) - acc + c.charCodeAt(0)) | 0, 0);
+                return ha - hb;
+            });
+            const randomSlice = shuffled.slice(0, randomCount);
+            const randomSet = new Set(randomSlice.map((m) => m.name));
+            // Insert random members at intervals through the clustered sequence.
+            const withRandom = [];
+            const interval = Math.max(2, Math.floor(clusteredMemories.length / randomCount));
+            let randomIdx = 0;
+            for (let i = 0; i < clusteredMemories.length; i++) {
+                const m = clusteredMemories[i];
+                if (m && !randomSet.has(m.name))
+                    withRandom.push(m);
+                if (i > 0 && i % interval === 0 && randomIdx < randomSlice.length) {
+                    const r = randomSlice[randomIdx++];
+                    if (r)
+                        withRandom.push(r);
+                }
+            }
+            // Append any remaining random members not yet inserted.
+            while (randomIdx < randomSlice.length) {
+                const r = randomSlice[randomIdx++];
+                if (r)
+                    withRandom.push(r);
+            }
+            finalClusteredMemories = withRandom;
+            warnings.push(`Anti-collapse: injected ${randomCount} random (non-similarity-driven) cluster member(s) into consolidation pool (fraction=${fraction}).`);
+        }
+    }
     const chunks = [];
-    for (let i = 0; i < clusteredMemories.length; i += chunkSize) {
-        chunks.push(clusteredMemories.slice(i, i + chunkSize));
+    for (let i = 0; i < finalClusteredMemories.length; i += chunkSize) {
+        chunks.push(finalClusteredMemories.slice(i, i + chunkSize));
     }
     // 2026-05-27 prompt-context fix: precompute body-hashes of pending
     // consolidate proposals once, so the per-chunk prompt can annotate
@@ -869,8 +895,46 @@ export async function akmConsolidate(opts = {}) {
     // 4h on this user's stack. See
     // /tmp/akm-health-investigations/tuning-reasons-investigation.md §Q3.
     const pendingProposalBodyHashes = loadPendingConsolidateProposalHashes(stashDir);
+    // ── Cold-start budget estimation ─────────────────────────────────────────────
+    // Estimate wall-clock cost BEFORE issuing any LLM calls. When a signal is
+    // provided and the estimated cost exceeds ~60% of the remaining budget we
+    // auto-reduce the pool and log the reduction so the run never starts work
+    // it cannot finish (avoiding SIGTERM mid-LLM-call).
+    //
+    // Formula: chunks.length × p90_chunk_seconds. The p90 comes from
+    // `opts.p90ChunkSecondsDefault` (caller-supplied, typically from the profile
+    // config); absent = 30 s (conservative default matching a medium local LLM).
+    //
+    // "Remaining budget" is read from a custom property on the AbortSignal if
+    // the caller (improve.ts) has attached one. Without it no auto-reduction
+    // fires but the check is still cheap to run.
+    if (chunks.length > 10 && opts.signal) {
+        const p90Chunk = opts.p90ChunkSecondsDefault ?? 30;
+        const estimatedSeconds = chunks.length * p90Chunk;
+        // remainingBudgetMs is a non-standard extension set by improve.ts when it
+        // creates the budget AbortController. Undefined = no budget information.
+        const budgetMs = opts.signal.remainingBudgetMs;
+        if (budgetMs !== undefined && budgetMs > 0) {
+            const remainingSeconds = budgetMs / 1000;
+            if (estimatedSeconds > remainingSeconds * 0.6) {
+                const safeCaps = Math.max(1, Math.floor((remainingSeconds * 0.6) / p90Chunk));
+                const removedChunks = chunks.length - safeCaps;
+                if (removedChunks > 0) {
+                    const msg = `[consolidate] cold-start budget: estimated ${estimatedSeconds.toFixed(0)}s > 60% of remaining ${remainingSeconds.toFixed(0)}s; ` +
+                        `reducing pool from ${chunks.length} to ${safeCaps} chunks (${removedChunks} deferred to next run).`;
+                    warn(msg);
+                    warnings.push(msg);
+                    chunks.splice(safeCaps);
+                }
+            }
+        }
+    }
     warn(`[consolidate] ${memories.length} memories / ${chunks.length} chunk(s) / chunk_size=${chunkSize}` +
         ` / pending-proposal hashes: ${pendingProposalBodyHashes.size}`);
+    // Consolidate output merges memories (non-wiki) → stash authoring standards.
+    // Resolved ONCE per run and passed to each chunk prompt (facts not re-read
+    // per chunk).
+    const standardsContext = resolveStashStandards(stashDir);
     const chunkOpsArrays = [];
     // Structured skip-reason histogram (2026-05-26): every deterministic
     // post-LLM op rejection site below also calls `pushSkipReason` so the
@@ -907,6 +971,13 @@ export async function akmConsolidate(opts = {}) {
     // judgedNoAction tracks memories the LLM saw inside a chunk but proposed
     // no op for. Computed per chunk as `chunk.length − unique(targetRefs in ops)`.
     let judgedNoAction = 0;
+    // Judged-state cache (#581): coarse outcome per memory NAME the LLM actually
+    // judged in a successfully-parsed chunk this run. "actioned" = an op targeted
+    // it; "no_action" = the LLM saw it and proposed nothing. Populated only when
+    // the cache is enabled (otherwise it stays empty and the post-loop recording
+    // step is a no-op). Memories in failed/aborted chunks are NOT recorded, so a
+    // transient LLM failure never poisons the cache into skipping them next run.
+    const judgedOutcomeByName = new Map();
     // 2026-05-27 cross-chunk double-count fix: refs that contributed to
     // judgedNoAction in their own chunk. When a different chunk's op references
     // one of these as a secondary and that op later fails, the ref would land
@@ -931,6 +1002,19 @@ export async function akmConsolidate(opts = {}) {
     const ABORT_MIN_CHUNKS = 4;
     const ABORT_FAILURE_RATE = 0.5;
     for (let chunkIdx = 0; chunkIdx < chunks.length; chunkIdx++) {
+        // Budget-signal check: break cleanly before the next LLM call if the
+        // caller's budget has been exhausted. Commits work done so far.
+        if (opts.signal?.aborted) {
+            const skipped = chunks.length - chunkIdx;
+            const msg = `[consolidate] budget signal aborted before chunk ${chunkIdx + 1}/${chunks.length}; ${skipped} chunk(s) not processed (partial_timeout — work done so far committed).`;
+            warn(msg);
+            warnings.push(msg);
+            // Account for memories in unprocessed chunks.
+            for (let i = chunkIdx; i < chunks.length; i++) {
+                failedChunkMemories += chunks[i].length;
+            }
+            break;
+        }
         // Abort if failure rate >= 50% over at least 4 processed chunks.
         if (totalChunksProcessed >= ABORT_MIN_CHUNKS) {
             const failureRate = totalChunksFailed / totalChunksProcessed;
@@ -968,7 +1052,7 @@ export async function akmConsolidate(opts = {}) {
             continue;
         }
         warn(`[consolidate] chunk ${chunkIdx + 1}/${chunks.length} (${chunk.length} memories) …`);
-        const userPrompt = buildChunkPrompt(sourceName, chunk, chunkIdx, chunks.length, bodyTruncation, pendingProposalBodyHashes);
+        const userPrompt = buildChunkPrompt(sourceName, chunk, chunkIdx, chunks.length, bodyTruncation, pendingProposalBodyHashes, standardsContext);
         let raw = await tryLlmFeature("memory_consolidation", config, async () => {
             if (!llmConfig)
                 return { ok: false, error: "No LLM configured for consolidation" };
@@ -1075,11 +1159,59 @@ export async function akmConsolidate(opts = {}) {
             if (!targetRefs.has(memRef)) {
                 chunkNoAction++;
                 judgedNoActionRefs.add(memRef);
+                // Judged-state cache (#581): the LLM saw this memory and proposed
+                // nothing → record judged-unchanged so the next run can skip it.
+                if (judgedCacheEnabled)
+                    judgedOutcomeByName.set(m.name, "no_action");
+            }
+            else if (judgedCacheEnabled) {
+                // An op targeted this memory → it was judged + actioned.
+                judgedOutcomeByName.set(m.name, "actioned");
             }
         }
         judgedNoAction += chunkNoAction;
         chunkOpsArrays.push(ops);
     }
+    // ── Judged-state cache recording (#581) ─────────────────────────────────────
+    // Persist judged state for every memory the LLM actually judged this run so
+    // the next run can skip the unchanged ones. Keyed by current content hash so
+    // a later body edit (different hash) re-enters the LLM pool. DEFAULT OFF and
+    // skipped under --dry-run (dry-run mutates nothing). Failed/aborted chunks
+    // contributed no entries to `judgedOutcomeByName`, so a transient LLM outage
+    // never caches a memory as judged.
+    if (judgedCacheEnabled && !opts.dryRun && judgedOutcomeByName.size > 0) {
+        // Use the shared state.db handle; open a local one as fallback.
+        const doRecord = (db) => {
+            const judgedAt = new Date(startMs).toISOString();
+            for (const [name, outcome] of judgedOutcomeByName) {
+                const hash = currentHashByName.get(name);
+                if (hash === undefined)
+                    continue;
+                upsertConsolidationJudged(db, {
+                    entryKey: `memory:${name}`,
+                    contentHash: hash,
+                    judgedAt,
+                    outcome,
+                });
+            }
+        };
+        if (sharedStateDb) {
+            try {
+                doRecord(sharedStateDb);
+            }
+            catch (e) {
+                warnings.push(`Judged-state cache: failed to record judged state: ${String(e)}`);
+            }
+        }
+        else {
+            try {
+                withStateDb((localDb) => doRecord(localDb));
+            }
+            catch (e) {
+                warnings.push(`Judged-state cache: failed to record judged state: ${String(e)}`);
+            }
+        }
+    }
     // Build the known-refs set from the already-filtered memory pool so
     // mergePlans() can reject LLM-hallucinated primary refs before execution.
     const knownRefs = new Set(memories.map((m) => `memory:${m.name}`));
@@ -1266,7 +1398,7 @@ export async function akmConsolidate(opts = {}) {
                 emitMergeFailureSkips(mergeResult.error);
                 continue;
             }
-            const mergedContent = mergeResult.content;
+            let mergedContent = mergeResult.content;
             // Validate frontmatter of merged content — must have a `---` block
             // with at minimum a `description` field. We parse via the hand-rolled
             // parser (cheap) AND require non-empty description. This guards against
@@ -1320,6 +1452,62 @@ export async function akmConsolidate(opts = {}) {
                 emitMergeFailureSkips("merge_participant_blocked");
                 continue;
             }
+            // WS-3b: Anti-collapse generation guard (step 8a).
+            // DEFAULT OFF. When antiCollapse.enabled, refuse to merge two assets both
+            // above generation N (default 2). This prevents the pipeline from
+            // building ever-deeper LLM-merged trees that lose the source fidelity
+            // of the original episodes.
+            const antiCollapseConfig = config.profiles?.improve?.default?.processes?.consolidate?.antiCollapse ??
+                {};
+            if (antiCollapseConfig.enabled) {
+                const allParticipants = [op.primary, ...op.secondaries];
+                const sourceGenerations = allParticipants.map((ref) => {
+                    const e = memoryByRef.get(ref);
+                    if (!e)
+                        return 0;
+                    try {
+                        const raw = fs.readFileSync(e.filePath, "utf8");
+                        const parsed = parseFrontmatter(raw);
+                        return readAssetGeneration(parsed.data);
+                    }
+                    catch {
+                        return 0;
+                    }
+                });
+                const generationCheck = checkGenerationGuard(sourceGenerations, antiCollapseConfig);
+                if (generationCheck.refused) {
+                    warnings.push(`Merge: ${generationCheck.reason}`);
+                    emitMergeFailureSkips("merge_generation_guard");
+                    continue;
+                }
+                // WS-3b: Lexical diversity check (step 8b).
+                // Low n-gram diversity ⇒ likely correlated-extraction artifact; raise merge threshold.
+                if (antiCollapseConfig.lexicalDiversityCheck !== false) {
+                    const bodies = allParticipants
+                        .map((ref) => {
+                        const e = memoryByRef.get(ref);
+                        if (!e)
+                            return "";
+                        try {
+                            const raw = fs.readFileSync(e.filePath, "utf8");
+                            return stripFrontmatterBody(raw);
+                        }
+                        catch {
+                            return "";
+                        }
+                    })
+                        .filter((b) => b.length > 0);
+                    const diversityCheck = checkLexicalDiversity(bodies, antiCollapseConfig);
+                    if (diversityCheck.lowDiversity) {
+                        // Low-diversity cluster: just warn (don't refuse merge since the dedup
+                        // path handles exact twins). The warning surfaces in health telemetry.
+                        warnings.push(`Merge: cluster around ${op.primary} has low lexical diversity (${diversityCheck.diversity?.toFixed(2) ?? "?"} < 0.30) — likely correlated extraction; merge proceeds but review is recommended.`);
+                    }
+                }
+                // Inject generation counter into merged content frontmatter (step 8a).
+                // merged.generation = max(sourceGenerations) + 1.
+                mergedContent = injectGenerationFrontmatter(mergedContent, sourceGenerations, allParticipants);
+            }
             // Backup secondaries before deleting
             for (const secRef of op.secondaries) {
                 const secEntry = memoryByRef.get(secRef);
@@ -1522,11 +1710,12 @@ export async function akmConsolidate(opts = {}) {
             //       is the load-bearing content, so dedup must hash on body only.
             //   (b) A prior run created a proposal for the same body under a
             //       different knowledgeRef slug.
-            const bodyHash = createHash("sha256").update(sourceBody, "utf8").digest("hex");
+            // Use cacheHash (case-preserving stripped body) to match the canonical
+            // hash domain used by the body-embedding cache and pending-proposal set.
+            const bodyHash = cacheHash(sourceBody);
             const allPendingConsolidateProposals = listProposals(stashDir, { status: "pending" }).filter((p) => p.source === "consolidate");
             const contentDupProposal = allPendingConsolidateProposals.find((p) => {
-                const otherBody = parseFrontmatter(p.payload.content).content.trim();
-                return createHash("sha256").update(otherBody, "utf8").digest("hex") === bodyHash;
+                return cacheHash(p.payload.content) === bodyHash;
             });
             if (contentDupProposal) {
                 warnings.push(`Skipping promote: identical body already pending as proposal ${contentDupProposal.id} (ref: ${contentDupProposal.ref}); skipping duplicate for ${op.ref} → ${knowledgeRef}`);
@@ -1608,6 +1797,18 @@ export async function akmConsolidate(opts = {}) {
             }
         }
         else if (op.op === "contradict") {
+            // Confidence gate: surface-level topic overlap causes false positives
+            // (investigation 2026-06-18). Require ≥0.92 confidence before writing
+            // contradiction edges. Missing confidence field defaults to 1.0 for
+            // backward compatibility with responses that predate this field.
+            const opConfidence = typeof op.confidence === "number"
+                ? op.confidence
+                : 1.0;
+            if (opConfidence < 0.92) {
+                warnings.push(`Contradict: confidence ${opConfidence.toFixed(2)} below 0.92 threshold for ${op.ref} <-> ${op.contradictedByRef} — skipping.`);
+                pushSkipReason("contradict", op.ref, "contradict_low_confidence");
+                continue;
+            }
             // C-3 / #382: Write contradictedBy edges so resolveFamilyContradictions
             // (the SCC resolver in memory-improve.ts) has edges to work on.
             // Zep arXiv:2501.13956 §3 — unified belief-revision with contradiction edges.
@@ -1647,37 +1848,19 @@ export async function akmConsolidate(opts = {}) {
         commitWriteTargetBoundary(target, `Consolidate: ${merged} merged, ${deleted} removed`);
     }
     cleanupJournal(stashDir, timestamp);
-    // TTL cleanup: remove archive entries older than archiveRetentionDays (default 90).
-    // C-5 / #391: emit an `archive_cleanup` event before each deletion so the
-    // audit trail records what was lost. Outbox pattern (EIP, Hohpe-Woolf) —
-    // any event that is recorded must be queryable; silent deletes are an anti-pattern.
-    const archiveDir = path.join(stashDir, ".akm", "archive");
-    if (fs.existsSync(archiveDir)) {
-        const retentionMs = (config.archiveRetentionDays ?? 90) * 86_400_000;
-        const cutoff = Date.now() - retentionMs;
-        for (const fname of fs.readdirSync(archiveDir)) {
-            const fp = path.join(archiveDir, fname);
-            try {
-                const stat = fs.statSync(fp);
-                if (stat.mtimeMs < cutoff) {
-                    // Emit event before deletion so the record survives the purge.
-                    appendEvent({
-                        eventType: "archive_cleanup",
-                        metadata: {
-                            file: fname,
-                            filePath: fp,
-                            ageMs: Date.now() - stat.mtimeMs,
-                            retentionMs,
-                        },
-                    });
-                    fs.unlinkSync(fp);
-                }
-            }
-            catch {
-                /* ignore race conditions */
-            }
-        }
-    }
+    // [signoff 2026-06-15] TTL archive cleanup machinery RETIRED (WS-3a).
+    // The elaborate archiveRetentionDays / archive-dir scan existed only to satisfy
+    // the old irrecoverability constraint. Stashes are now git-backed, so git
+    // history is the recovery path — no bespoke archive TTL needed. Any files in
+    // .akm/archive/ will stay there harmlessly until the operator prunes them with
+    // `git rm` or `find .akm/archive -mtime +90 -delete`. Changed N files this
+    // run; recover any via `git show <sha>:<path>` or `git restore <path>`.
+    if (merged > 0 || deleted > 0 || dedupCollapsed > 0) {
+        const totalChanged = merged + deleted + dedupCollapsed;
+        warnings.push(`Changed ${totalChanged} file(s) this run. Recover any via git if needed (git history is the backstop).`);
+    }
+    const runDurationMs = Date.now() - startMs;
+    const budgetFraction = opts.runBudgetMs !== undefined && opts.runBudgetMs > 0 ? runDurationMs / opts.runBudgetMs : undefined;
     return {
         schemaVersion: 1,
         ok: true,
@@ -1687,7 +1870,10 @@ export async function akmConsolidate(opts = {}) {
         target: sourceName,
         processed: memories.length,
         merged,
-        deleted,
+        // #617: fold the deterministic dedup pre-pass collapses into the reported
+        // deleted count. Each collapse removed exactly one variant file with NO
+        // LLM call before the LLM pass ran on the pruned pool.
+        deleted: deleted + dedupCollapsed,
         promoted,
         contradicted,
         failedChunks: totalChunksFailed,
@@ -1697,235 +1883,19 @@ export async function akmConsolidate(opts = {}) {
         mergedSecondaries,
         failedChunkMemories,
         warnings,
-        durationMs: Date.now() - startMs,
+        durationMs: runDurationMs,
+        perfTelemetry: {
+            dedupPoolSize: perfMs.dedupPoolSize,
+            llmPoolSize,
+            judgedCacheSkipped: perfMs.judgedCacheSkipped,
+            embedMs: embedTelemetry.embedMs,
+            embedCacheHits: embedTelemetry.cacheHits,
+            embedCacheMisses: embedTelemetry.cacheMisses,
+            ...(budgetFraction !== undefined ? { estimatedBudgetFractionUsed: budgetFraction } : {}),
+        },
     };
 }
 // ── Helpers ─────────────────────────────────────────────────────────────────
-// ── LLM-output sanitization ─────────────────────────────────────────────────
-//
-// Three classes of LLM defect have been observed across hundreds of
-// consolidate proposals (see audit notes in this branch):
-//
-//   1. Code-fence leakage: the entire merged asset is wrapped in
-//      ```markdown … ``` (or ```yaml … ```) despite the prompt forbidding
-//      fences. The post-processor used to pass this through verbatim, so the
-//      first character of the asset content became a backtick rather than
-//      `---`, defeating the frontmatter parser.
-//   2. YAML quote-escaping bugs: descriptions like `'"Specialty intro...:`
-//      with unbalanced quotes that break the YAML reader. The post-processor
-//      historically passed the LLM's raw scalar straight into a manually
-//      assembled `description: <raw>` line.
-//   3. Truncated descriptions hitting token cutoffs — the model's max_tokens
-//      runs out mid-sentence, leaving things like
-//      `description: "Tables in narrow column containers need max-width:100% +"`
-//      with no closing context.
-//
-// `sanitizeMergedContent` and `validateProposalFrontmatter` defend against
-// all three at the point where LLM output is consumed.
-/**
- * Attempt to recover a frontmatter block that is missing its closing `---`.
- *
- * Scans lines after the opening `---` for the first blank line or the first
- * line that cannot be a YAML scalar (i.e. not a key-value, indented
- * continuation, comment, or list item). Injects `---` before that line so
- * the normal parser can proceed.
- *
- * Returns the patched string on success, or `null` if the structure is too
- * ambiguous to recover safely (e.g. no opening `---`, or no body content
- * found after the frontmatter key-value lines).
- */
-function recoverMalformedFrontmatter(raw) {
-    if (!raw.startsWith("---"))
-        return null;
-    const lines = raw.split(/\r?\n/);
-    // Skip the opening `---` line (index 0).
-    let insertAt = -1;
-    for (let i = 1; i < lines.length; i++) {
-        const line = lines[i];
-        // A blank line marks the end of the frontmatter block in many YAML variants.
-        if (line.trim() === "") {
-            insertAt = i;
-            break;
-        }
-        // A line that is clearly body content: doesn't look like a YAML key, an
-        // indented continuation, a comment, or a sequence item.
-        const isYaml = /^\w[\w-]*\s*:/.test(line) || // key: value
-            /^\s+\S/.test(line) || // indented continuation / nested
-            /^\s*#/.test(line) || // YAML comment
-            /^\s*-\s/.test(line); // sequence item
-        if (!isYaml) {
-            insertAt = i;
-            break;
-        }
-    }
-    if (insertAt < 0)
-        return null;
-    const result = [...lines.slice(0, insertAt), "---", ...lines.slice(insertAt)].join("\n");
-    return result;
-}
-/**
- * Outer-fence stripper specific to consolidate. Unlike the shared
- * `stripMarkdownFences` helper (which only handles markdown fences), this
- * variant additionally recognises `yaml` and bare-language fences and refuses
- * to strip an unbalanced fence — i.e. a leading ``` with no trailing ``` is
- * treated as a malformed response, not partially sanitized.
- *
- * Returns `null` when only one half of a fence pair is present (caller
- * should reject the response entirely).
- */
-export function stripOuterCodeFence(raw) {
-    const trimmed = raw.trim();
-    const leading = trimmed.match(/^```(?:markdown|md|yaml|yml)?\s*\r?\n/i);
-    const trailing = trimmed.match(/\r?\n```\s*$/);
-    if (!leading && !trailing)
-        return { content: trimmed, stripped: false };
-    if (!leading || !trailing)
-        return null; // unbalanced — refuse
-    const inner = trimmed.slice(leading[0].length, trimmed.length - trailing[0].length).trim();
-    return { content: inner, stripped: true };
-}
-export function sanitizeMergedContent(raw) {
-    // Step 1: Strip outer code fence.
-    // Recovery path: if only the leading fence is present, strip it and continue
-    // provided the inner content starts with `---`. Trailing-only fences are NOT
-    // recovered — a trailing ``` is more likely a body code block than a forgotten
-    // wrapper, so recovering would silently corrupt the body.
-    let body;
-    {
-        const fenceResult = stripOuterCodeFence(raw);
-        if (fenceResult) {
-            body = fenceResult.content;
-        }
-        else {
-            const trimmed = raw.trim();
-            const leadingMatch = trimmed.match(/^```(?:markdown|md|yaml|yml)?\s*\r?\n([\s\S]*)$/i);
-            const inner = leadingMatch ? leadingMatch[1].trim() : null;
-            if (!inner?.startsWith("---")) {
-                return { ok: false, reason: "UNBALANCED_CODE_FENCE" };
-            }
-            body = inner;
-        }
-    }
-    // Strip <think> blocks (some local models still emit them despite system prompts).
-    body = body.replace(/<think>[\s\S]*?<\/think>/gi, "").trim();
-    // Step 2: Verify frontmatter sentinel.
-    // Recovery path: LLM sometimes emits 1-2 lines of preamble (e.g. "Here is the
-    // merged content:") before the `---`. Accept if `---` appears within 300 chars.
-    // Beyond that it's more likely a body section divider, not a frontmatter start.
-    if (!body.startsWith("---")) {
-        const nlIdx = body.indexOf("\n---");
-        if (nlIdx >= 0 && nlIdx < 300) {
-            body = body.slice(nlIdx + 1);
-        }
-        else {
-            return { ok: false, reason: "MISSING_FRONTMATTER_SENTINEL" };
-        }
-    }
-    // Extract frontmatter block.
-    // Recovery path: LLM sometimes omits the closing `---` delimiter. Detect this
-    // by scanning lines after the opening `---` for the first blank line or the
-    // first line that isn't a YAML key-value pair, then inject `---` there.
-    let match = body.match(/^---\r?\n([\s\S]*?)\r?\n---(?:\r\n|\r|\n|$)([\s\S]*)$/);
-    if (!match) {
-        const recovered = recoverMalformedFrontmatter(body);
-        if (recovered) {
-            match = recovered.match(/^---\r?\n([\s\S]*?)\r?\n---(?:\r\n|\r|\n|$)([\s\S]*)$/);
-        }
-        if (!match) {
-            return { ok: false, reason: "MALFORMED_FRONTMATTER_BLOCK" };
-        }
-    }
-    // Re-parse via the yaml library so any quote-escaping mistakes either get
-    // normalised or surface as a parse error we can reject.
-    // Recovery: if the strict yaml library fails, fall back to the lenient
-    // hand-rolled parseFrontmatter parser, which tolerates common LLM YAML
-    // quirks (unescaped special chars, bare scalars, etc.). If it recovers
-    // at least one key, proceed — serializeFrontmatter below will re-serialize
-    // cleanly. Only reject if both parsers fail to extract any data.
-    let parsedFm;
-    try {
-        parsedFm = yamlParse(match[1]);
-    }
-    catch (e) {
-        const fallback = parseFrontmatter(`---\n${match[1]}\n---\n${match[2]}`);
-        if (fallback.frontmatter !== null && Object.keys(fallback.data).length > 0) {
-            parsedFm = fallback.data;
-        }
-        else {
-            return { ok: false, reason: `INVALID_YAML: ${e instanceof Error ? e.message : String(e)}` };
-        }
-    }
-    if (parsedFm === null || typeof parsedFm !== "object" || Array.isArray(parsedFm)) {
-        return { ok: false, reason: "FRONTMATTER_NOT_OBJECT" };
-    }
-    const fm = parsedFm;
-    // Normalise placeholder leaks like `updated: today`, `updated: {today: null}`,
-    // `updated: now`, etc. The consolidate prompt instructs the LLM not to emit
-    // these, but small models still do. Replace any such leak with today's ISO
-    // date OR drop the field if we can't safely normalise it.
-    normalizeUpdatedField(fm);
-    // Re-serialise via yaml.stringify to fix any quoting quirks.
-    let serialized;
-    try {
-        serialized = serializeFrontmatter(fm);
-    }
-    catch (e) {
-        return { ok: false, reason: `YAML_STRINGIFY_FAILED: ${e instanceof Error ? e.message : String(e)}` };
-    }
-    const cleaned = assembleAssetFromString(serialized, match[2]);
-    return { ok: true, result: { content: cleaned, frontmatter: fm } };
-}
-/**
- * Mutate `fm.updated` in place to normalise placeholder leaks emitted by the
- * LLM. The consolidate prompt forbids these, but small models still produce
- * literal `today` / `{today: null}` / `now` values.
- *
- * Rules:
- *   - A real ISO-style date string (YYYY-MM-DD, optionally with time) stays as-is.
- *   - A Date object (some YAML parsers materialise dates) is converted to its
- *     ISO yyyy-mm-dd form.
- *   - A placeholder string ("today", "now", "{today}", "${today}", template
- *     variables) is replaced with today's ISO date.
- *   - A map/object (e.g. `{today: null}`) is replaced with today's ISO date.
- *   - `null`, empty string, missing → left alone (no field added; reviewers
- *     should not silently gain metadata they didn't write).
- *
- * Exported for unit testing.
- */
-export function normalizeUpdatedField(fm) {
-    if (!("updated" in fm))
-        return;
-    const v = fm.updated;
-    if (v === null || v === undefined || v === "")
-        return;
-    const todayIso = new Date().toISOString().slice(0, 10);
-    if (v instanceof Date) {
-        fm.updated = v.toISOString().slice(0, 10);
-        return;
-    }
-    if (typeof v === "string") {
-        const trimmed = v.trim().toLowerCase();
-        if (/^\d{4}-\d{2}-\d{2}/.test(v.trim()))
-            return; // already a real date
-        if (trimmed === "today" ||
-            trimmed === "now" ||
-            trimmed === "{today}" ||
-            // biome-ignore lint/suspicious/noTemplateCurlyInString: matches the literal user-typed placeholder text "${today}" so we can normalize it to today's ISO date
-            trimmed === "${today}" ||
-            trimmed === "{{today}}" ||
-            /^\{?\s*today\s*\}?$/.test(trimmed)) {
-            fm.updated = todayIso;
-            return;
-        }
-        // Unknown string format — leave alone so it's visible in the diff.
-        return;
-    }
-    if (typeof v === "object") {
-        // Maps like `{today: null}`, `{now: null}` — clearly a template leak.
-        fm.updated = todayIso;
-        return;
-    }
-}
 /**
  * Normalise a knowledge slug for variant-aware deduplication. Collapses:
  *   - date suffixes (`-may-2026`, `-2026-05-03`, `-2026`)
@@ -2004,7 +1974,7 @@ function parseSinceToIso(since) {
     const multiplier = { m: 60_000, h: 3_600_000, d: 86_400_000 }[m[2]];
     return new Date(Date.now() - parseInt(m[1], 10) * multiplier).toISOString();
 }
-export function narrowToIncrementalCandidates(memories, since, warnings) {
+export function narrowToIncrementalCandidates(memories, since, warnings, neighborsPerChanged = 5) {
     const sinceIso = parseSinceToIso(since);
     const isChanged = (m) => {
         try {
@@ -2019,7 +1989,6 @@ export function narrowToIncrementalCandidates(memories, since, warnings) {
         return [];
     if (changed.length === memories.length)
         return memories;
-    const NEIGHBORS_PER_CHANGED = 5;
     const byName = new Map(memories.map((m) => [m.name, m]));
     const keep = new Set(changed.map((m) => m.name));
     let db;
@@ -2029,7 +1998,7 @@ export function narrowToIncrementalCandidates(memories, since, warnings) {
             const id = findEntryIdByRef(db, `memory:${m.name}`);
             if (id === undefined)
                 continue;
-            for (const hit of getNeighborsByEntryId(db, id, NEIGHBORS_PER_CHANGED + 1)) {
+            for (const hit of getNeighborsByEntryId(db, id, neighborsPerChanged + 1)) {
                 if (hit.id === id)
                     continue;
                 const entry = getEntryById(db, hit.id);
@@ -2067,6 +2036,10 @@ function loadMemoriesForSource(source, stashDir, warnings) {
             return path.resolve(e.stashDir) === path.resolve(source);
         })
             .filter((e) => isConsolidationEligibleMemoryName(e.entry.name))
+            // #632 — exclude session-capture telemetry (checkpoints) from the
+            // consolidation pool, mirroring recombine. Their bodies are pipeline
+            // bookkeeping, so consolidating them burns LLM calls on pure noise.
+            .filter((e) => !isSessionCaptureMemoryName(e.entry.name))
             // Skip stale DB entries whose file was deleted by a prior run but not yet
             // re-indexed. Without this guard the deleted file's ref appears in chunks
             // sent to the LLM, which then proposes a second delete → delete_failed
@@ -2100,6 +2073,8 @@ function loadMemoriesForSource(source, stashDir, warnings) {
                 const name = fname.replace(/\.md$/, "");
                 if (!isConsolidationEligibleMemoryName(name))
                     continue;
+                if (isSessionCaptureMemoryName(name))
+                    continue; // #632 — skip telemetry checkpoints
                 memories.push({ name, filePath, description: "", tags: [], stashDir: fsStashDir });
             }
         }
@@ -2231,8 +2206,8 @@ async function generateMergedContent(config, primaryRef, primaryBody, secondaryR
             }
             normalizeUpdatedField(repairedFmData);
             const repairedYaml = serializeFrontmatter(repairedFmData);
-            const bodyPart = mergedFm.content ?? "";
-            return { content: `---\n${repairedYaml}\n---\n${bodyPart}` };
+            const bodyPart = typeof mergedFm.content === "string" ? mergedFm.content : "";
+            return { content: assembleAssetFromString(repairedYaml, bodyPart) };
         }
     }
     catch {