session 2026-07-19 night: one-door authorized (REVIEWED-66) + V1 built + @2 governing (PENDING-66 closed) + audit_cruft kindle_residue

claude/memory/skill-harvest-register.md

@@ -558,3 +558,11 @@ No new skill proposal. Two register notes: **(1)** the S1 proposal **implementat
 | **re-anchor = re-verify: reconstruct the old bound state by sha-match** | verification-ladder entry | When re-anchoring any sha-bound artifact after an upstream edit: reconstruct the OLD bound state from git by matching the declared sha (`git rev-list` the file; hash each historical blob), prove the old→live delta is EXACTLY the expected hunks, and map every changed line against the declared spans/anchors BEFORE re-stamping any sha. Stored-text anchors verify ellipsis-aware (split on `...`/`…`, fragments in order) — a declared condensation divergence then surfaces as the index's own exception, not a false failure. Proven 2026-07-19 on all five after-the-reply sources + the 48-anchor reading-index the completeness grep caught. Will recur constantly under the one-door reconversion posture (every Route-B graduation is a re-anchor). | `reference-verification-ladder.md` | **PROPOSED** |
 
 **Reinforcement (no new proposal):** the 2026-06-18 `/wake-up` telos-patch proposal ("surface the *why* when the thread touches the engine's purpose") gains a second evidence instance — the 2026-07-19 full-map answer under-carried the shortlist/Making telos lane and the steward had to correct it twice (drift-pattern `map-drawn-from-production-lanes-not-telos`). Still awaiting steward judgment with the rest of the standing review.
+
+### Harvest 2026-07-19 (night — one-door authorized · V1 built · @2 governing)
+
+| Element | Kind | One-line | Where it lands | Status |
+|---|---|---|---|---|
+| **implement-the-relation-not-an-approximation** | verification-ladder entry | When code implements a RULED relation/contract (an equivalence relation, a gate criterion), the acceptance path must BE the relation — any structural approximation (word-aligned matching of a string-defined relation; a sampled check of a total property) is a silent SECOND relation that under- or over-implements the ruled one and drifts independently. Earned 2026-07-19: V1's stage-2 word-aligned matcher was fine under @1 but silently under-implemented @2 (whose guillemet rewrite folds across word boundaries) — caught only because the @2 test-flip ran red; acceptance is now normalized-string containment, the relation itself, with the word machinery demoted to evidence-location. **Sub-note (enumerated-charset writes):** when a spec/code artifact carries an ENUMERATED character class, write it as explicit escapes and verify by set-equality against the stated list — an invisible-literal (exotic spaces, combining marks) can silently diverge from its own enumeration at write time (self-caught same session; the Q2 condition's exact failure shape). | `reference-verification-ladder.md` | **PROPOSED** |
+
+*Reinforcements recorded, not re-proposed:* **census-before-pattern / read-the-substrate** earned a sharp instance (a `\b`-anchored kindle_residue regex would have missed EVERY real glued token — PASS-BUT-FALSELY at birth; reading Eichmann's actual anchors first caught it). **checkable-claim** ran three number-reconciliations in one session (72=67+5 œ/Œ · 4,712=2×2,356 tokens/links · 1,305=1,297+8 md/txt) — counting, never harmonizing. **implementation-is-a-second-gate** (thrice-earned, still PROPOSED above) gained a fourth instance: the @2 test-flip IS the instrument — ratified text exercised by having to act on it surfaced two contract-semantics findings (the approximation; fragment-as-quotable-sub-span).