@gamaze/hicortex 0.22.2 → 0.23.0

package/README.md

@@ -91,8 +91,8 @@ Bare top-level keys (`"serverUrl": …` at the root of `openclaw.json`) still wo
 | When | What | How |
 |------|------|-----|
 | Agent start | Standing identity (`## Identity`) + recent lessons fetched fresh and injected | CC SessionStart hook (calls `hicortex learnings-identity`; `lessons-context` kept as an alias so existing installed hooks don't break) / Hermes plugin `system_prompt_block` / OC `before_agent_start` hook |
-| Every prompt (0.14) | A compact **recall index** of relevant memories is injected — one line per memory; the agent lazy-loads full content with `hicortex_get` only when needed | All five harnesses call server `POST /recall-index` per turn: CC UserPromptSubmit hook (`hicortex recall-hook`), Hermes plugin `prefetch` (0.7.0; falls back to `/search` injection against a pre-0.14 server), OC `before_agent_start` hook (fires per inbound message), Pi extension `before_agent_start` (0.20), opencode plugin messages-transform hook (0.21). Turn-based dedup per session; resets on new session/compaction. Fail-soft |
-| Nightly | Denoise sessions → POST /distill → server distills + embeds + stores → consolidate (score, reflect, link, decay) | Automatic pipeline — no manual steps |
+| Every prompt (0.14) | A compact **recall index** of relevant memories is injected — one line per memory; the agent lazy-loads full content with `hicortex_get` only when needed | All five harnesses call server `POST /recall-index` per turn: CC UserPromptSubmit hook (`hicortex recall-hook`), Hermes plugin `prefetch` (0.7.0; falls back to `/search` injection against a pre-0.14 server), OC `before_agent_start` hook (fires per inbound message), Pi extension `before_agent_start` (0.20), opencode plugin messages-transform hook (0.21). Turn-based dedup per session; resets on new session/compaction. Fail-soft. Harness plumbing (task-notifications, command envelopes) is skipped deterministically at BOTH the hook and the server — a wrapper carrying real prose still recalls on the full text |
+| Nightly | Denoise sessions → POST /distill → server distills + embeds + stores → consolidate (score, reflect, link, decay) | Automatic pipeline — no manual steps. Volatile status entries (GitHub ticket states, version bumps, commit/push states) are dropped at a deterministic entry gate beside the substance gate — every drop rides the `/distill` response's `dropped` audit trail; decision/policy wording ("switched from X to Y", "never relicense…") always escapes the gate |
 
 **Exposure vs use (0.14):** appearing in the recall index only marks a memory as *shown* (it stops decaying while topically active); fetching it with `hicortex_get` marks it as *used* (durable strengthening). Memory importance is driven by what agents actually use, not by what was pushed at them.
 
@@ -216,6 +216,8 @@ npx @gamaze/hicortex classify-domains          # Backfill domain tags over the c
 npx @gamaze/hicortex classify-types            # Reclassify memory types (episode/fact/decision)
 npx @gamaze/hicortex dedup                     # Preview near-duplicate memory clusters (dry run, no changes)
 npx @gamaze/hicortex dedup --apply             # Merge near-duplicate clusters (backs up the DB first)
+npx @gamaze/hicortex sweep-volatile            # Preview volatile GitHub-status/version rows the gate would retire (dry run)
+npx @gamaze/hicortex sweep-volatile --apply    # Retire them via absorb (backs up the DB first; audit trail; never deletes)
 npx @gamaze/hicortex identity show [name]       # Print the standing identity layer (see Identity Layer)
 npx @gamaze/hicortex identity edit <name>       # Edit an identity section in $EDITOR
 npx @gamaze/hicortex identity show --agent <id> # Show a specific agent's resolved identity (0.13)

package/assets/dashboard.html

@@ -39,7 +39,7 @@
     - Overview: the counter/bar/headroom key off LIVE memories (absorbed
       rows show as an "absorbed (kept as evidence)" note); the counter is
       STATIC (no replay tracking — owner ruling 2026-09-13).
-    - Recall card: the RAW uses-per-showing ratio + a 7v7-nights trend, NO
+    - Recall card: the RAW uses-per-showing ratio, NO
       grade verdict (#426 owner semantics ruling: recall depends on the
       conversation, higher is not a target).
     - Rail header: "Fleet".
@@ -89,7 +89,7 @@
       default, which used to degrade the corner to a standalone Upgrade
       button. The upgrade CTA is a MENU ITEM (same plan-unset eligibility
       the shipped #365 code used); no standalone upgrade button remains.
-    - RECALL CARD: below the band at most ONE subtle line (trend glyph +
+    - RECALL CARD: below the band at most ONE subtle line (zone word +
       muted "provisional"; pill chrome dropped); the zone word rides the
       marker, the zone-label row is gone — the card stops setting the row
       height.
@@ -135,7 +135,7 @@
       DOWN into the siblings' chart slot (header → big → band → foot).
       "provisional" is hover-only now — a tooltip on the band (the card
       stays marked; the visible word failed to communicate). Vocabulary
-      amendment: Low / HEALTHY / Overfetching. Marker + trend stay.
+      amendment: Low / HEALTHY / Overfetching. Marker + zone word stay.
 
   REPLAY + RANGE (#452, owner directives 2026-09-16):
     - 30-DAY DEFAULT FOR ALL VIEWS: the replay ledger, the band and the
@@ -397,10 +397,6 @@
   .big .n.noconf{ font-size:14.5px; font-weight:520; letter-spacing:0; color:var(--dim); }
   html:not([data-theme="light"]) .big .n.noconf{ text-shadow:none; }
   .big .l{ font-size:10px; color:var(--faint); text-transform:uppercase; letter-spacing:.13em; }
-  .trend{ font-size:12px; line-height:1; flex:none; }
-  .trend.up{ color:var(--green-mid); }
-  .trend.down{ color:var(--red); }
-  .trend.flat{ color:var(--dim); }
   /* THE card sub-line — one text type */
   .csub{ font-size:11px; color:var(--dim); white-space:nowrap; overflow:hidden; text-overflow:ellipsis; }
   .csub b{ color:var(--green-mid); font-weight:600; }
@@ -419,7 +415,7 @@
      row). The visible "provisional" word is GONE — the marking is hover-only
      now, a tooltip on the band (the visible word failed to communicate).
      Vocabulary (owner 2026-09-14): Low / Healthy / Overfetching. The raw
-     number stays tooltip-only on the marker; the trend glyph stays (mock
+     number stays tooltip-only on the marker (mock
      family). Colors: the band + word carry concrete per-theme colors from
      the JS stop table (setTheme re-renders), the dot rides theme tokens. */
   .rec-word.none{ color:var(--dim); }
@@ -3178,23 +3174,6 @@ function buildTokNights(used, cap, periodStart){
   }
 }
 
-/* trend = mean of the last 7 nights vs the prior 7 (a DEFINED window — the
-   old first-vs-last-over-the-range comparison had no window at all). Titles
-   state the window so the arrow is never mysterious. */
-function recallTrend(){
-  const flat={cls:'flat',sym:'▬'};
-  if(!D||!D.series) return Object.assign(flat,{title:'no snapshot data'});
-  const vals=D.series.map(s=>s.metrics&&s.metrics.adoption?s.metrics.adoption.uses_per_showing:null).filter(v=>typeof v==='number');
-  const last7=vals.slice(-7), prior7=vals.slice(-14,-7);
-  const mean=a=>a.length?a.reduce((x,y)=>x+y,0)/a.length:null;
-  const a=mean(prior7), b=mean(last7);
-  if(a===null||b===null) return Object.assign(flat,{title:'trend needs 7+ nights of history'});
-  const d=b-a;
-  if(d>0.005) return {cls:'up',sym:'↗',title:'up — mean of the last 7 nights vs the prior 7 ('+b.toFixed(3)+' vs '+a.toFixed(3)+')'};
-  if(d<-0.005) return {cls:'down',sym:'&#8600;',title:'down — mean of the last 7 nights vs the prior 7 ('+b.toFixed(3)+' vs '+a.toFixed(3)+')'};
-  return Object.assign(flat,{title:'flat — mean of the last 7 nights vs the prior 7 ('+b.toFixed(3)+' vs '+a.toFixed(3)+')'});
-}
-
 /* ---------- Recall level — the designed U-band (#426 final ruling
    2026-09-13). A continuous red→green→red gradient spans the PROVISIONAL
    zone edges (echoed by /dashboard/field as thresholds.recall_uses — the
@@ -3291,16 +3270,11 @@ function renderRecBand(ups){
     word.style.color=''; word.style.textShadow='none';
     word.title='no showings yet';
   }
-  /* trend arrow — the mock family's 7v7 window, unchanged */
-  if(typeof ups==='number'){
-    const tr=recallTrend();
-    el('rec-trend').className='trend '+tr.cls;
-    el('rec-trend').innerHTML=tr.sym;
-    el('rec-trend').title=tr.title;
-  }else{
-    el('rec-trend').className='trend flat'; el('rec-trend').innerHTML='&#9644;';
-    el('rec-trend').title='no showings yet';
-  }
+  /* No trend arrow (owner 2026-09-20: removed — unclear). Its renderer and
+     the element writes are removed WITH it: a leftover write to a removed
+     element is a boot-killing TypeError whenever the recall band carries
+     real data (bedrock 0.22.2: "0 memories" — boot died before the counter
+     was ever seeded). */
 }
 
 /* ---------- Memory Precision — Level 1, the pushed index (#476). A compact

package/dist/calibration.d.ts

@@ -117,11 +117,11 @@ export declare const MEMORY_PRECISION_WINDOW_DAYS = 90;
 /**
  * Cosine at/above which a pushed line reads "redundant" (restating the
  * session's standing context — the top lessons + identity every session-start
- * hook injects). Direction-only display guidance, never a gate. 0.80 mirrors
- * SUPERSESSION_MIN_SIMILARITY: at/above it a pair reads as the same
- * statement, below it as topical overlap — a reasonable first anchor until
- * the judged-calibration follow-up measures the real distribution (the
- * #426 provisional-anchor posture).
+ * hook injects). Direction-only display guidance, never a gate. 0.80 — the
+ * anchor it was seeded from (the retired supersession stage's floor, #206-B)
+ * at/above which a pair reads as the same statement, below it as topical
+ * overlap — a reasonable first anchor until the judged-calibration follow-up
+ * measures the real distribution (the #426 provisional-anchor posture).
  */
 export declare const MEMORY_PRECISION_REDUNDANT_ABOVE = 0.8;
 /**
@@ -138,6 +138,23 @@ export declare const MEMORY_PRECISION_DIVERGENCE_MIN_SHOWN = 5;
  * the same trust boundary as /distill's denoised conversation text).
  */
 export declare const MEMORY_PRECISION_PROMPT_EXCERPT_CHARS = 256;
+/**
+ * Kill-switch for the deterministic volatility gate (#489, owner decision 5):
+ * false disables isVolatileStatusEntry's use in the distill path (entries
+ * keep flowing as before) AND the sweep's candidacy rule — one release-managed
+ * constant, no config knob (#408 discipline: moving it is a release decision,
+ * never a runtime one).
+ */
+export declare const VOLATILE_STATUS_FILTER = true;
+/**
+ * Max entry length the volatility gate will classify as volatile (#489) — the
+ * operationalization of the spec's "DOMINANT content matches": distiller
+ * bullets are one-liners, so a status shape in a SHORT entry dominates it,
+ * while the same clause riding inside long mixed prose does not (kept — the
+ * substance gate's precision-over-recall bias, applied to a second gate).
+ * 300 chars ≈ 3× the longest shape in the #489 KEEP/DROP corpus.
+ */
+export declare const VOLATILE_GATE_MAX_CHARS = 300;
 /** Semantic-similarity share of the composite score. 0.50 (raised from 0.40
  *  in the 0.15.2 rebalance, #191 Phase B): on the production corpus effective
  *  strength (0.30) outweighed what similarity could recover — hardened old
@@ -273,13 +290,11 @@ export declare const IMPORTANCE_SETTLE_WINDOW_DAYS = 3;
  *  get the one verdict call. 0.92 — measured on the #191 mechanical audit
  *  corpus (89 clusters / 110 excess rows; data/audit-20260729). */
 export declare const DEDUP_AUTO_MERGE_THRESHOLD = 0.92;
-/** Minimum cosine for a nightly supersession candidate pair (#100 stage,
- *  0.15.0): one classify-tier call per pair above the bar. */
-export declare const SUPERSESSION_MIN_SIMILARITY = 0.8;
 /** Minimum cosine for a reconsolidation correction pair (#384). Deliberately
- *  wider than supersession's 0.80: a retraction often rides inside an
- *  otherwise unrelated memory; the verdict + confidence gate carry the
- *  precision. */
+ *  wide: a retraction often rides inside an otherwise unrelated memory; the
+ *  verdict + confidence gate carry the precision. (The retired Stage 3.7
+ *  supersession scan's 0.80 floor — #206-B, owner decision 6 — is subsumed:
+ *  3.8 scans all shapes from this wider floor.) */
 export declare const CORRECTION_MIN_SIMILARITY = 0.75;
 /** Minimum verdict confidence for the REWRITE (and #392 merge-apply) fork
  *  (#384): below it a `corrects` degrades to mark-only — a weak mark is

package/dist/calibration.js

@@ -24,8 +24,8 @@
  * the removal is never silent.
  */
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.STAGE_TRUTH_STRENGTH = exports.STAGE_BELIEF_STRENGTH = exports.STAGE_FADING_STRENGTH = exports.STAGE_FADING_DAYS = exports.WEAK_PRIMARY_FLOOR = exports.CORRECTION_REWRITE_MIN_CONFIDENCE = exports.CORRECTION_MIN_SIMILARITY = exports.SUPERSESSION_MIN_SIMILARITY = exports.DEDUP_AUTO_MERGE_THRESHOLD = exports.IMPORTANCE_SETTLE_WINDOW_DAYS = exports.IMPORTANCE_CEILING = exports.BM25_WEIGHT_DOMAIN = exports.BM25_WEIGHT_PROJECT = exports.BM25_WEIGHT_BODY = exports.BOTH_CHANNEL_BOOST = exports.RRF_VECTOR_WEIGHT = exports.RRF_FTS_WEIGHT = exports.RRF_COMPOSITE_WEIGHT = exports.RRF_K = exports.SCOPE_AFFINITY_WEIGHT = exports.SUPERSEDED_DEMOTION = exports.RECENCY_HEAD_DAYS = exports.RECENCY_HEAD_WEIGHT = exports.RECENCY_HOURLY_DECAY = exports.SCORE_RECENCY_WEIGHT = exports.CONNECTIONS_SATURATION_DEGREE = exports.SCORE_CONNECTIONS_WEIGHT = exports.SCORE_STRENGTH_WEIGHT = exports.SCORE_SIMILARITY_WEIGHT = exports.MEMORY_PRECISION_PROMPT_EXCERPT_CHARS = exports.MEMORY_PRECISION_DIVERGENCE_MIN_SHOWN = exports.MEMORY_PRECISION_REDUNDANT_ABOVE = exports.MEMORY_PRECISION_WINDOW_DAYS = exports.RECALL_USES_AXIS_MAX = exports.RECALL_USES_NORMAL_MAX = exports.RECALL_USES_LOW_MAX = exports.RECALL_RESHOW_TURNS = exports.NOVELTY_FLOOR_SLOTS = exports.RECALL_TITLE_CHARS = exports.RECALL_MIN_PROMPT_CHARS = exports.RECALL_MAX_ITEMS = exports.RECALL_MIN_SIMILARITY = exports.SESSION_INTENT_WEIGHT = exports.COLD_EXPOSURE_SLOTS = exports.RECENT_WINDOW_DAYS = exports.RECENT_LIMIT = exports.SEARCH_LIMIT = exports.PROMOTION_STRENGTH_FLOOR = exports.PROMOTION_RATE = exports.DECAY_HALF_LIFE_DAYS = void 0;
-exports.DIAGNOSTIC_ENV_TIER = exports.OLLAMA_FLUSH_WAIT_MS = exports.OLLAMA_FLUSH_EVERY = exports.NUM_CTX = exports.DEFAULT_FIRST_RUN_LOOKBACK_DAYS = exports.ENRICH_STRENGTH_DELTA = void 0;
+exports.STAGE_BELIEF_STRENGTH = exports.STAGE_FADING_STRENGTH = exports.STAGE_FADING_DAYS = exports.WEAK_PRIMARY_FLOOR = exports.CORRECTION_REWRITE_MIN_CONFIDENCE = exports.CORRECTION_MIN_SIMILARITY = exports.DEDUP_AUTO_MERGE_THRESHOLD = exports.IMPORTANCE_SETTLE_WINDOW_DAYS = exports.IMPORTANCE_CEILING = exports.BM25_WEIGHT_DOMAIN = exports.BM25_WEIGHT_PROJECT = exports.BM25_WEIGHT_BODY = exports.BOTH_CHANNEL_BOOST = exports.RRF_VECTOR_WEIGHT = exports.RRF_FTS_WEIGHT = exports.RRF_COMPOSITE_WEIGHT = exports.RRF_K = exports.SCOPE_AFFINITY_WEIGHT = exports.SUPERSEDED_DEMOTION = exports.RECENCY_HEAD_DAYS = exports.RECENCY_HEAD_WEIGHT = exports.RECENCY_HOURLY_DECAY = exports.SCORE_RECENCY_WEIGHT = exports.CONNECTIONS_SATURATION_DEGREE = exports.SCORE_CONNECTIONS_WEIGHT = exports.SCORE_STRENGTH_WEIGHT = exports.SCORE_SIMILARITY_WEIGHT = exports.VOLATILE_GATE_MAX_CHARS = exports.VOLATILE_STATUS_FILTER = exports.MEMORY_PRECISION_PROMPT_EXCERPT_CHARS = exports.MEMORY_PRECISION_DIVERGENCE_MIN_SHOWN = exports.MEMORY_PRECISION_REDUNDANT_ABOVE = exports.MEMORY_PRECISION_WINDOW_DAYS = exports.RECALL_USES_AXIS_MAX = exports.RECALL_USES_NORMAL_MAX = exports.RECALL_USES_LOW_MAX = exports.RECALL_RESHOW_TURNS = exports.NOVELTY_FLOOR_SLOTS = exports.RECALL_TITLE_CHARS = exports.RECALL_MIN_PROMPT_CHARS = exports.RECALL_MAX_ITEMS = exports.RECALL_MIN_SIMILARITY = exports.SESSION_INTENT_WEIGHT = exports.COLD_EXPOSURE_SLOTS = exports.RECENT_WINDOW_DAYS = exports.RECENT_LIMIT = exports.SEARCH_LIMIT = exports.PROMOTION_STRENGTH_FLOOR = exports.PROMOTION_RATE = exports.DECAY_HALF_LIFE_DAYS = void 0;
+exports.DIAGNOSTIC_ENV_TIER = exports.OLLAMA_FLUSH_WAIT_MS = exports.OLLAMA_FLUSH_EVERY = exports.NUM_CTX = exports.DEFAULT_FIRST_RUN_LOOKBACK_DAYS = exports.ENRICH_STRENGTH_DELTA = exports.STAGE_TRUTH_STRENGTH = void 0;
 exports.resolveNumCtx = resolveNumCtx;
 exports.resolveOllamaFlushEvery = resolveOllamaFlushEvery;
 exports.resolveOllamaFlushWaitMs = resolveOllamaFlushWaitMs;
@@ -144,11 +144,11 @@ exports.MEMORY_PRECISION_WINDOW_DAYS = 90;
 /**
  * Cosine at/above which a pushed line reads "redundant" (restating the
  * session's standing context — the top lessons + identity every session-start
- * hook injects). Direction-only display guidance, never a gate. 0.80 mirrors
- * SUPERSESSION_MIN_SIMILARITY: at/above it a pair reads as the same
- * statement, below it as topical overlap — a reasonable first anchor until
- * the judged-calibration follow-up measures the real distribution (the
- * #426 provisional-anchor posture).
+ * hook injects). Direction-only display guidance, never a gate. 0.80 — the
+ * anchor it was seeded from (the retired supersession stage's floor, #206-B)
+ * at/above which a pair reads as the same statement, below it as topical
+ * overlap — a reasonable first anchor until the judged-calibration follow-up
+ * measures the real distribution (the #426 provisional-anchor posture).
  */
 exports.MEMORY_PRECISION_REDUNDANT_ABOVE = 0.8;
 /**
@@ -166,6 +166,28 @@ exports.MEMORY_PRECISION_DIVERGENCE_MIN_SHOWN = 5;
  */
 exports.MEMORY_PRECISION_PROMPT_EXCERPT_CHARS = 256;
 // ---------------------------------------------------------------------------
+// Volatility gate (#489) — deterministic drop of GH-ticket-status /
+// version-bump / commit-state entries at the distill gate, and the one-shot
+// store sweep that reuses the same gate.
+// ---------------------------------------------------------------------------
+/**
+ * Kill-switch for the deterministic volatility gate (#489, owner decision 5):
+ * false disables isVolatileStatusEntry's use in the distill path (entries
+ * keep flowing as before) AND the sweep's candidacy rule — one release-managed
+ * constant, no config knob (#408 discipline: moving it is a release decision,
+ * never a runtime one).
+ */
+exports.VOLATILE_STATUS_FILTER = true;
+/**
+ * Max entry length the volatility gate will classify as volatile (#489) — the
+ * operationalization of the spec's "DOMINANT content matches": distiller
+ * bullets are one-liners, so a status shape in a SHORT entry dominates it,
+ * while the same clause riding inside long mixed prose does not (kept — the
+ * substance gate's precision-over-recall bias, applied to a second gate).
+ * 300 chars ≈ 3× the longest shape in the #489 KEEP/DROP corpus.
+ */
+exports.VOLATILE_GATE_MAX_CHARS = 300;
+// ---------------------------------------------------------------------------
 // Composite ranking weights (was: score*Weight, supersededDemotion,
 // *AffinityWeight, rrf*; the pre-#430 freshness bonus constants merged into
 // the RECENCY family below)
@@ -315,13 +337,11 @@ exports.IMPORTANCE_SETTLE_WINDOW_DAYS = 3;
  *  get the one verdict call. 0.92 — measured on the #191 mechanical audit
  *  corpus (89 clusters / 110 excess rows; data/audit-20260729). */
 exports.DEDUP_AUTO_MERGE_THRESHOLD = 0.92;
-/** Minimum cosine for a nightly supersession candidate pair (#100 stage,
- *  0.15.0): one classify-tier call per pair above the bar. */
-exports.SUPERSESSION_MIN_SIMILARITY = 0.80;
 /** Minimum cosine for a reconsolidation correction pair (#384). Deliberately
- *  wider than supersession's 0.80: a retraction often rides inside an
- *  otherwise unrelated memory; the verdict + confidence gate carry the
- *  precision. */
+ *  wide: a retraction often rides inside an otherwise unrelated memory; the
+ *  verdict + confidence gate carry the precision. (The retired Stage 3.7
+ *  supersession scan's 0.80 floor — #206-B, owner decision 6 — is subsumed:
+ *  3.8 scans all shapes from this wider floor.) */
 exports.CORRECTION_MIN_SIMILARITY = 0.75;
 /** Minimum verdict confidence for the REWRITE (and #392 merge-apply) fork
  *  (#384): below it a `corrects` degrades to mark-only — a weak mark is

package/dist/capture.js

@@ -303,7 +303,10 @@ async function captureBatches(batches, opts) {
                     lastConfirmedEnd = seg.segEnd;
                 console.log(`[hicortex]     → ${result.distilled ?? 0} memories (segment ${body.segment_id})`);
                 for (const d of result.dropped ?? []) {
-                    console.log(`[hicortex]     Substance gate: dropped "${d}"`);
+                    // #489: the response array carries BOTH gate kinds (substance +
+                    // volatility) — the server's own log names the exact gate; this
+                    // client-side line is the same #156 trail, labeled generically.
+                    console.log(`[hicortex]     Distill gate: dropped "${d}"`);
                 }
             }
             else if (result.status === 200) {

package/dist/cli.d.ts

@@ -16,6 +16,9 @@
  *              dedup --apply           Execute the merge (default: dry run);
  *                                      losers are absorbed (kept as evidence,
  *                                      hidden from recall), not deleted
+ *   sweep-volatile  One-shot sweep of volatile GH-status/version rows (#489)
+ *              sweep-volatile --apply  Demote matches via absorb (default:
+ *                                      dry run; backup first, audit trail)
  *   history    Show a memory's rewrite history, or roll one back (issue #384)
  *              history --rollback <row>  Undo one rewrite + un-absorb triggers
  *   status     Show config, DB stats, adapter status

package/dist/cli.js

@@ -17,6 +17,9 @@
  *              dedup --apply           Execute the merge (default: dry run);
  *                                      losers are absorbed (kept as evidence,
  *                                      hidden from recall), not deleted
+ *   sweep-volatile  One-shot sweep of volatile GH-status/version rows (#489)
+ *              sweep-volatile --apply  Demote matches via absorb (default:
+ *                                      dry run; backup first, audit trail)
  *   history    Show a memory's rewrite history, or roll one back (issue #384)
  *              history --rollback <row>  Undo one rewrite + un-absorb triggers
  *   status     Show config, DB stats, adapter status
@@ -306,6 +309,29 @@ switch (command) {
         });
         break;
     }
+    case "sweep-volatile": {
+        // One-shot store sweep of volatile GH-status/version rows (#489,
+        // deliverable iii): reuses the distill path's deterministic gate over the
+        // stored corpus. Mirrors `dedup`: dry-run default, --apply explicit
+        // (backup first, absorb — never delete, volatile_sweep_log audit trail).
+        const args = process.argv.slice(3);
+        let dbPath;
+        try {
+            dbPath = (0, cli_args_js_1.readValueFlag)(args, "--db");
+        }
+        catch {
+            console.error("[hicortex] sweep-volatile: --db requires a path value");
+            process.exit(1);
+        }
+        const sweepOptions = { apply: args.includes("--apply"), dbPath };
+        import("./sweep-volatile.js").then(({ runSweepVolatile }) => {
+            runSweepVolatile(sweepOptions).catch((err) => {
+                console.error(err instanceof Error ? err.message : `[hicortex] sweep-volatile failed: ${err}`);
+                process.exit(1);
+            });
+        });
+        break;
+    }
     case "history": {
         // Rewrite history: audit + one-command rollback (#384). Listing is
         // read-only; --rollback mutates (restores prior content/status and

package/dist/cluster.d.ts

@@ -25,8 +25,12 @@ export declare class UnionFind {
 export declare function clusterEdges(edges: Edge[], threshold: number): string[][];
 /** Excess = sum(cluster size − 1) — rows that would disappear if every cluster merged to one. */
 export declare function clusterExcess(clusters: string[][]): number;
-/** Metadata fields a merge candidate cluster must agree on. `privacy` is on the row (the
- *  column still exists) but is NOT a merge-safety field — vestigial since 0.16.2. */
+/** Metadata row shape for the merge-safety check. `privacy` and `source_agent`
+ *  are on the row (the columns still exist) but are NOT merge-safety fields:
+ *  privacy is vestigial since 0.16.2, and the agent rail was REMOVED by owner
+ *  decision (#206 decision 2, 2026-09-21) — cross-agent clusters merge, because
+ *  attribution survives on the retained absorbed-loser row and no recall path
+ *  filters by source_agent. Project is the only rail. */
 export interface ClusterMetaRow {
     project: string | null;
     privacy: string | null;
@@ -34,10 +38,9 @@ export interface ClusterMetaRow {
 }
 export interface ClusterMetadataMismatch {
     projectMismatch: boolean;
-    sourceAgentMismatch: boolean;
 }
-/** Do cluster members disagree on project / source_agent? (merge-safety input for #100).
- *  Privacy is intentionally NOT checked — it is vestigial since 0.16.2. */
+/** Do cluster members disagree on project? (merge-safety input for #100/#206.)
+ *  Privacy and source_agent are intentionally NOT checked — see ClusterMetaRow. */
 export declare function clusterMetadataMismatch(members: ClusterMetaRow[]): ClusterMetadataMismatch;
 /**
  * Build the max-cosine edge set via top-K KNN on `memory_vectors`, keeping

package/dist/cluster.js

@@ -69,15 +69,11 @@ function clusterEdges(edges, threshold) {
 function clusterExcess(clusters) {
     return clusters.reduce((sum, c) => sum + (c.length - 1), 0);
 }
-/** Do cluster members disagree on project / source_agent? (merge-safety input for #100).
- *  Privacy is intentionally NOT checked — it is vestigial since 0.16.2. */
+/** Do cluster members disagree on project? (merge-safety input for #100/#206.)
+ *  Privacy and source_agent are intentionally NOT checked — see ClusterMetaRow. */
 function clusterMetadataMismatch(members) {
     const projects = new Set(members.map((m) => m.project ?? "\u0000null"));
-    const agents = new Set(members.map((m) => m.source_agent));
-    return {
-        projectMismatch: projects.size > 1,
-        sourceAgentMismatch: agents.size > 1,
-    };
+    return { projectMismatch: projects.size > 1 };
 }
 // ---------------------------------------------------------------------------
 // KNN edge building

package/dist/consolidate.d.ts

@@ -257,65 +257,6 @@ export declare function classifyLinkCandidates(candidates: LinkCandidate[]): Pro
  * boundary from the l2ToCosine calibration is preserved.
  */
 export declare function classifyRelationship(source: Memory, target: Memory, similarity: number): string;
-/** Default minimum COSINE similarity for a supersession candidate pair —
- *  RELEASE-MANAGED since #408 (calibration.ts SUPERSESSION_MIN_SIMILARITY). */
-export declare const DEFAULT_SUPERSESSION_MIN_SIMILARITY = 0.8;
-export interface SupersessionOptions {
-    /** Candidate-pair cosine floor. Release-managed default (calibration.ts);
-     *  this field is the eval/test seam. Invalid → default. */
-    minSimilarity?: number;
-    /** The run-wide pipeline deadline (#405) — checked at each candidate
-     *  boundary; on expiry the scan stops and the cursor holds at the last
-     *  fully-considered candidate (resumed next run). */
-    deadline?: RunDeadline;
-}
-export interface SupersessionStageResult {
-    scanned: number;
-    evaluated: number;
-    superseded: number;
-    skipped_infra: number;
-    skipped_idempotent: number;
-    cursor: number;
-}
-/**
- * Build the constrained supersession-check prompt. Content is truncated the
- * same width as domain-classify.ts's classifier (1500 chars) — this is a
- * classify-tier call with the same cost profile.
- */
-export declare function buildSupersessionPrompt(oldContent: string, newContent: string): string;
-/**
- * Parse the model's supersession verdict. Returns the boolean on a valid
- * reply, or null on anything unparseable (caller skips the pair — no retry,
- * unlike domain-classify's tag classifier; a missed pair is retried naturally
- * when this stage revisits the corpus).
- */
-export declare function parseSupersessionReply(reply: string): boolean | null;
-/**
- * Nightly supersession-detection stage. Scans memories/rowid > cursor whose
- * shape is supersedable (decision/correction/fact/state — isSupersedableShape),
- * checks each against its older same-shape neighbors, and links confirmed
- * supersessions. Dry-run performs discovery + the free idempotency check only —
- * no LLM calls, no writes, no
- * cursor persistence (mirrors stageImportance/stageContentDomains's dry-run
- * convention of never spending budget on a preview).
- *
- * Cursor discipline is DELIBERATELY simple (owner amendment): the cursor
- * advances past a candidate once its neighbor set has been considered,
- * REGARDLESS of whether every pair got an LLM call (call budget) or a clean
- * verdict (infra skip) — missing one pair is acceptable and self-heals next
- * time this memory's neighborhood is re-examined via a NEWER memory's own
- * candidacy. It only stops SHORT of a candidate when the budget is already
- * exhausted before that candidate starts, so the cursor never skips a
- * candidate that was never looked at.
- *
- * #405: the cursor persists after EVERY fully-considered candidate (the
- * post-#404 reconsolidation pattern), not at stage end — a run killed or
- * deadline-deferred mid-stage loses at most the candidate in flight. No
- * orphan clamp is needed (unlike reconsolidation): supersession applies each
- * verdict's link immediately, so `cursor = candidate.__rowid` always sits
- * after all of that candidate's writes.
- */
-export declare function stageSupersession(db: Database.Database, llm: LlmClient, budget: BudgetTracker, embedFn: EmbedFn, dryRun: boolean, stateDir: string | undefined, options?: SupersessionOptions): Promise<SupersessionStageResult>;
 /**
  * Exported for the #191 eval baseline (src/eval/decay-eval.ts) so the audit
  * runs the REAL production prune predicate against a DB snapshot instead of
@@ -404,7 +345,7 @@ export interface DomainStageOptions {
      */
     weakPrimaryFloor?: number;
 }
-export declare function runConsolidation(db: Database.Database, llm: LlmClient, embedFn: EmbedFn, dryRun?: boolean, skipReflection?: boolean, stateDir?: string, domainOptions?: DomainStageOptions, supersessionOptions?: SupersessionOptions, 
+export declare function runConsolidation(db: Database.Database, llm: LlmClient, embedFn: EmbedFn, dryRun?: boolean, skipReflection?: boolean, stateDir?: string, domainOptions?: DomainStageOptions, 
 /** The ONE per-run LLM-call ceiling (#405/#241). The caller resolves
  *  `nightlyLlmCallBudget` from config (consolidateMaxLlmCalls is a
  *  deprecated alias — resolveNightlyLlmCallBudget) and passes it; unset →
@@ -416,10 +357,9 @@ budgetMaxCalls?: number,
 memorySoftCap?: number, 
 /** Reconsolidation-stage knobs (#384) — eval/test seams since #408 (the
  *  values are release-managed calibration constants; nightly.ts threads
- *  NOTHING), exactly like supersessionOptions above; unset fields → the
- *  stage's calibration defaults. Appended AFTER the pre-#384 params so
- *  every existing positional caller (tests, hosted nightly) keeps its
- *  argument meaning. */
+ *  NOTHING); unset fields → the stage's calibration defaults. The
+ *  pre-#206-B supersessionOptions param that sat here is retired with its
+ *  stage — callers updated in the same change. */
 reconsolidationOptions?: ReconsolidationOptions, 
 /**
  * The run-wide pipeline deadline (#405), created at nightly start and