session 2026-07-07: MemPalace wind-down decided (evidenced audit) + KG exported to files

Diagnosed the "upgrade" as a 300-behind bge-m3 fork, not a standard install; the
morning "self-healed" drift was 108 chronic quarantines (data intact). Independent
46-transcript audit: search+KG not load-bearing (value migrated to files). Steward
decided (evidenced) to wind down palace-memory; execute fresh. KG exported to
knowledge-graph.jsonl (329 triples + 437 entities). Wind-down plan + BMF-lessons
report outline captured. Demote-on-promote; skill harvest appended.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dsqe1x23NgWaRdiFWSotur

claude/memory/skill-harvest-register.md

@@ -361,3 +361,13 @@ The single place proposed skills live so they don't evaporate between sessions.
 - **PROMOTE `completion-is-a-tripwire` → Symmetria §3 flag (recurrence met; the watch-item from 2026-07-06 evening).** Tonight IS the recurrence the watch-item awaited — and it arrived in a sharper, *objective* form: a **verification SUITE whose every check is orthogonal to a failure class passes green while claiming to verify.** The 5 `verify_conversion.py` gates (cruft / structure / word-final-& / U+FFFD / not-shattered) all passed Olson while sparse word-level OCR error (`Ihave`, garbled Greek) sat untouched — none of the 5 can *see* it, and a joined word even *improves* `word_sanity`. The wall of green read as thoroughness; "graduated + gates PASS" oversold a converted-tier V-SCAN text; the steward caught it. **Proposed flag text:** *a passing gate-suite is not coverage — ask which failure class each check can actually see before trusting the green; the feeling of "done" after a fluent pass or a wall of green is the cue to verify the tail, not the signal to ship.* Kin to `rank-on-fields-you-actually-write` (a mechanism that can't do what it claims, silently) + `census-through-truncation`. Home: Symmetria §3 — fold the objective "gate-suite blind to the failure class" form into the subjective `completion-is-a-tripwire`. **PROPOSED — recurrence-promoted per the governed pattern, not minted-from-one.**
 - **Not a skill (tracked elsewhere):** the **V-SCAN OCR-vs-scan verification stage + model-plurality doctrine** is a chamber `[PROPOSAL]` (parked; steward-sequenced priority once the REVIEWED-49 folding completes) — captured in the session file + ledger + KG, becomes its own PENDING when we return. Design invariant: verifier FLAGS discrepancies, never autonomously corrects; olmOCR = local localizer / executor-vision = adjudication-assist / human = apex; first step = the olmOCR reformed-spelling probe.
 - **CLAUDE.md freshness (mechanical, applied):** chamber-library CLAUDE.md spec-version pointer bumped v2.0.0 → v2.0.1 at the A2 landing — a tracker-style pointer update applied at wrap (not a discipline/posture change).
+
+### Harvest 2026-07-07 (evening — MemPalace wind-down decided)
+
+| Element | Kind | One-line | Where it lands | Status |
+|---|---|---|---|---|
+| **quote a long-job ETA only from an observed rate, never model-size intuition** | verification-ladder entry | Twice today I gave a re-embed ETA from gut ("15–45 min") and was wrong by ~30×; the steward caught it. The fix was **measuring** — rows-per-elapsed on the running job (2000/27min) + a 150-chunk benchmark → **~21 h**, grounded and confirmed. Rule: any long job (re-embed, migration, bulk convert) ETA must come from a **measured throughput sample** (a short timed run or the live job's own rate), stated with the sample. "It's a small/fast model" is not an estimate. Kin to measure-toolchain-before-spec / measure-the-font's-true-advance. | `reference-verification-ladder.md` | **PROPOSED** |
+| **Symmetria §3 flag: reassuring-verb-before-verifying-the-mechanism** | Symmetria §3 flag | Reaching for a comforting characterization ("self-healed", "fine", "recovered", "handled") *before* verifying the actual mechanism is contamination shape — it smooths the steward instead of informing them. Caught hard today: I called the morning HNSW quarantine "self-healed"; verifying showed **108 chronic quarantines** — damage-control, not repair. Antidote: name the mechanism you verified, not the outcome you hope for; if unchecked, say "I don't yet know if it recovered." Kin to `causal-story-before-reading-render`, one level up (the *valence* of the claim, not just its causal story). | Symmetria §3 | **PROPOSED** |
+| **`mempalace-diagnose` — RETIRE the proposal** | retire | AUTHORIZED 2026-06-05 (build-on-need). Now **decorative**: the steward decided (evidenced) to **wind down** palace-memory MemPalace. Building a MemPalace diagnostic skill for a substrate we're leaving is the inverse of proportionate. Retire on decommission. (The reusable, substrate-agnostic instruments from tonight — measure-ETA, from-sqlite-rebuilds-only-chroma, verify-install-isn't-standard-before-upgrade — are captured as the ladder/flag entries above, not tied to MemPalace.) | skill-harvest register | **PROPOSED (retire)** |
+
+*(Three, all earned in the MemPalace session. The first two are substrate-agnostic instruments that outlive MemPalace; the third is honest housekeeping — don't build tooling for what we're decommissioning. None autonomous. Also fulfilled-not-proposed: the standing `/wrap-up §5` "mirror KG to files → palace disposable" proposal was **realized** this session as the `knowledge-graph.jsonl` export — mark it done when the decommission wires wake to read it.)*