session 2026-07-11: footnote recognizer arc — hash-locality ratified (12405a5), split-anchor REVIEWED-54a (a6828ac), census re-audit, generalization REVIEWED-54b pt1 block-id+url-decode (ef0cee0); session file + ledger + KG(+4) + skill-harvest

claude/memory/skill-harvest-register.md

@@ -402,3 +402,15 @@ The single place proposed skills live so they don't evaporate between sessions.
 *Mechanical (applied at wrap per §1.6 CLAUDE.md-freshness): chamber `CLAUDE.md` fleet line gains the two footnote-pre-processor tools (tools entered the fleet — event-based update). Discipline-level items (the CLAUDE.md:67 re-anchor note from 2026-07-09; a §V ratified-spec cross-ref for the pre-processor) remain surfaced-not-applied.*
 
 *(One ladder proposal, load-bearing — it's a clean generalization the jurist forced, and the guard it names has already caught a real defect. The session was otherwise disciplined [returns held: measure-before-build, verify-against-substrate, honest-refuse-built-into-the-tool]; no other harvest manufactured.)*
+
+### Harvest 2026-07-11 (Sennett split-anchor → census re-audit → recognizer generalization, three jurist loops)
+
+| Element | Kind | One-line | Where it lands | Status |
+|---|---|---|---|---|
+| **re-audit coverage with the TOOL's recognizer, not the classifier that shares its blind spot** | verification-ladder entry | When a classifier and the tool it feeds share a predicate, the classifier's *mis*-classifications masquerade as genuine "new categories." To size a remaining gap, re-audit by running the **tool's own recognizer** over the corpus — not the census classifier. Earned 2026-07-11 (jurist's census-misclassification flag): Sennett was census-filed "one-directional" yet was covered by the split-anchor recognizer; running the current recognizer over all 252 sources collapsed the apparent "5 remaining families" to **one reciprocal mirror with the id in 3 positions**. Cheaper to re-measure than to build a widen for a phantom family. Kin to *measure-don't-speculate* / *censused-routes*, specialized to shared-blind-spot classifiers. | `reference-verification-ladder.md` | **PROPOSED** |
+| **extending a tool re-tests its foundations** | verification-ladder note / feedback | Building an extension exercises shared machinery the original's tests never hit — so a widen's `--validate` should assert the *shared* invariants too, not only the new predicate. Earned 2026-07-11: building the block-id handler surfaced a **latent duplicate-id defect in the already-shipped Sennett code** (`<p id=X … id=X>`, pandoc-tolerated, undetected by the prior fixture). The extension's dup-id assertion is what caught it. | `reference-verification-ladder.md` (or a feedback memory) | **PROPOSED** |
+| **classify a change by MECHANISM, not by how big it feels** | feedback memory OR Symmetria §3 note | Reflexively labeled the recognizer generalization "PROPOSAL" because it *felt* large; the jurist's own FIX/PROPOSAL test (does it touch the **verification** — the guard, a new mechanism, or pandoc's prose role?) showed most of it is **FIX**. Over-classifying to appear cautious is the *safe* direction (over-governance, not a contamination risk), but it still burdens the loop and mis-frames the ask; classify by mechanism. Milder cousin of the executor-bias flags. | feedback memory or Symmetria §3 | **PROPOSED (low-urgency; safe-direction error)** |
+
+*Mechanical (applied at wrap per §1.6 CLAUDE.md-freshness): chamber `CLAUDE.md` fleet line + `conversion-runbook.yaml` policy updated for REVIEWED-54a (split-anchor) + REVIEWED-54b pt1 (block-id + URL-decode) — tools' recognizer classes changed (event-based). §V ratified-spec cross-ref for the pre-processor still surfaced-not-applied (discipline-level).*
+
+*(Three proposals, none manufactured: the first two are concrete instruments the session proved [the re-audit collapsed a phantom taxonomy; the extension caught a real latent defect]; the third is an honest recalibration, flagged as a safe-direction error so it isn't over-weighted. REVIEWED-54b part 2 [inner-anchor + letter-marker] will earn its own review after build.)*