feat(jme): Phase 4 — auto-prune low-confidence facts when vector ceiling is active

[PiotrCoderDroid]

JME Phase 4 — Auto-prune on ceiling

Problema

JME utilidad en day-14 (2026-07-27): 13.3% global (gate: ≥39%). Hipótesis: a medida que jme_facts crece, facts marginales (baja confianza, desactualizados) llenan el LIMIT 500 del query window en queryMemory(), desplazando facts útiles y degradando el recall.

Phase 3 ya medía el ceiling y emitía un warn. Phase 4 actúa sobre él.

Cambios

src/memory/jme.ts

• Nueva constante PHASE4_CONFIDENCE_FLOOR = 0.5: umbral más agresivo que el baseline 0.4 de pruneExpiredFacts(). Solo se usa cuando el ceiling está activo.
• ConsolidateResult.factsPruned: number: nuevo campo para observabilidad.
• Auto-prune en consolidateAll(): cuando embeddingCount >= VECTOR_CEILING_WARN (400), ejecuta DELETE (expired + confidence < 0.5 + ts > 30d ago) antes de la extracción Haiku. Conteo en result.factsPruned.

src/rituals/scheduler.ts

• Log de jme-consolidate incluye , N p4-pruned cuando Phase 4 actuó (omitido cuando factsPruned == 0).

src/memory/jme.test.ts

• Test 1: inserta exactamente 400 facts con embedding + baja confianza → factsPruned >= 400, fact de alta confianza sobrevive.
• Test 2: 399 facts (bajo ceiling) → factsPruned == 0.

Invariantes

• pruneExpiredFacts() standalone no cambia (threshold 0.4 intacto).
• El auto-prune corre antes del Haiku call — no toca el path de extracción.
• Idempotente: si no hay elegibles, factsPruned = 0, log sin cambio.

Tests

• jme.test.ts: 36/36 ✅ (2 tests nuevos de Phase 4)
• Typecheck: PASS ✅
• Nota: 2 fallos pre-existentes en scheduler.test.ts (pm-daily-rebalance) — verificados en main sin mis cambios, no relacionados con este PR.

[kosm1x]

Closing without merging. The change is inert against live data — every one of its three conditions matches an empty set today, so merging it would add a destructive DELETE path that can never fire for the reason it was written.

Measured on data/mc.db at the time of review (129 jme_facts rows, all embedded):

condition	live reality
fires when embeddingCount >= VECTOR_CEILING_WARN (400)	129 — ~34 days away at the current ~8 facts/day
confidence < PHASE4_CONFIDENCE_FLOOR (0.5)	0 rows — the minimum confidence in the table is 0.72
ts < 30 days ago	0 rows — the oldest fact is 16 days old

All 10 rows the predicate would remove come from the expires_at clause, which already existed. Raising the floor 0.4 → 0.5 changes the outcome for exactly zero facts, and the extractor has never emitted a confidence below 0.72 across the whole table.

Three specific problems, independent of the above:

1. It forks a function that already exists and is never called. pruneExpiredFacts() (src/memory/jme.ts:528) contains the byte-identical DELETE, differing only in the constant. It has no caller anywhere in production. This PR copy-pastes its SQL inline rather than wiring it up or parameterizing the floor, leaving two identical destructive statements over the same table to keep in sync — one of which nobody runs.

2. The tests manufacture the population that reality lacks. They seed 400 facts at confidence 0.3 aged 31 days — precisely the shape production does not contain. 36/36 green while proving nothing about whether the code ever executes. A test whose fixture supplies the missing precondition cannot discriminate between "works" and "unreachable."

3. The branch head is not what was built. Local worktree is at 95258b3 (based on current main); this PR's head is 0e1f898 on an older lineage — same message, different SHA, 25 ahead / 1 behind. And the commit message says "Closes JME Phase 4 (pending post-day-14 veredicto decision)" — closed and pending at once.

If the vector ceiling becomes real (~early September at the current rate), the honest fix is to call the pruneExpiredFacts() that already exists and pick a floor from the observed confidence distribution — not to fork it with a threshold below the minimum value the extractor has ever produced.

Commit preserved as 95258b3 / 0e1f898; nothing is lost by closing.

[kosm1x]

Closed per operator review — see the analysis above. Not merged; no code from this branch reaches main.