fix(doctor,storage): a broken AI feature nobody was told about (D13, D15)

CHANGELOG.md

@@ -255,6 +255,37 @@ All notable changes to MeMesh are documented here.
   append path gets the same content guard, so this dedup is now a property of
   writing an observation at all, not just of these three hooks.
 
+- **`memesh doctor` now says when an AI-backed feature has quietly stopped
+  working.** `llm_telemetry` recorded every Smart-Mode call but nothing read
+  it for a health signal — measured on a real graph, `guard_proposer` had
+  failed all 69 of its calls, every one, across the four days it ran between
+  2026-08-28 and 2026-09-01, invisible to doctor. A new zero-cost check
+  (`inspectLlmTelemetryHealth`, no `--probe`
+  gate) reads the last 7 days and warns when a flow's recent calls (≥3) all
+  failed; a shorter window than `memesh telemetry`'s 30-day default is
+  deliberate — a longer one blends in stale successes and hides a flow that
+  has been 100% broken all week. Silent (not a false PASS) when the table is
+  empty or a flow has never run; informational when calls exist and none are
+  100%-failing. New `doctor.msg.llm-telemetry.silent-failure` catalogue entry
+  in all 11 dashboard locales.
+
+- **A split lesson no longer starts its life carrying its parent bucket's
+  stale recall history.** `splitFusedLessons` moves a fused `-other` bucket's
+  observations into their own entities but left `recall_hits`/`recall_misses`
+  on the now-empty, archived bucket forever — measured, 4 of 49 already-split
+  shells on a real graph still carried nonzero history (one at 3 hits / 61
+  misses), which `scripts/audit/measure-signals.mjs`'s unfiltered
+  `SUM(recall_hits) FROM entities` counts to this day. The bucket's rate is
+  itself an artifact of the fusion being repaired — a bucket fusing many
+  unrelated lessons gets injected broadly and cited rarely, which is not a
+  measurement of what any one lesson in it is worth — so it is not divided or
+  copied onto the successors (they correctly start at 0/0); the shell's own
+  counters are zeroed and the number is kept for forensic reading under
+  `metadata.retired_recall`. A new one-shot repair
+  (`repairFusedLessonShellHistory`) retires the history on shells already in
+  this state, and `scripts/audit/memory-invariants.mjs` gained
+  `split-lesson-shell-carries-no-recall-history` to keep it caught.
+
 - **The Ollama host guard now rebuilds the request origin instead of forwarding
   the configured string.** `resolveOllamaHost` used to validate a persisted
   `llm.host` and then pass the same string to `fetch`, which left CodeQL alert

dist/storage/graph-repairs.d.ts

@@ -4,6 +4,7 @@ export declare const ZERO_EDIT_RETRACT_KEY = "session_zero_edit_retract";
 export declare const FUSED_LESSON_SPLIT_KEY = "fused_lesson_split";
 export declare const ARCHIVED_FTS_ROWS_KEY = "archived_fts_rows";
 export declare const ARCHIVED_VECTOR_ROWS_KEY = "archived_vector_rows";
+export declare const FUSED_LESSON_SHELL_HISTORY_RESET_KEY = "fused_lesson_shell_history_reset";
 export declare function dedupeObservations(db: MemeshDatabase): number;
 export declare function bashWritesFiles(command: string): boolean;
 export declare function retractZeroEditClaims(db: MemeshDatabase): number;
@@ -15,4 +16,5 @@ export declare function dropArchivedIndexRows(db: MemeshDatabase): {
     ftsRows: number;
     vectorRows: number;
 };
+export declare function repairFusedLessonShellHistory(db: MemeshDatabase): number;
 //# sourceMappingURL=graph-repairs.d.ts.map