session 2026-06-28: Musil graduated + Position I complete + gold-standard gate fixed (verify RAW_HTML) + 2 policies (engine-sources-graduate, notes-are-the-work) + maturation dialogue

claude/memory/skill-harvest-register.md

@@ -257,3 +257,15 @@ The single place proposed skills live so they don't evaporate between sessions.
 | **`feedback-resurface-banked-notes-before-rederiving`** | DONE (memory, not skill) | Created this session — read the banked note before re-deriving; re-derivation drifts (page-numbers mis-recalled as provenance; Levi gate re-raised 3×). The engine's purpose turned on my practice. | BUILT (memory) |
 | verification-ladder note: **OCR word-guard passes on SCRAMBLED text** | ladder entry (proposed) | A verbatim word-guard that checks word PRESENCE cannot catch reading-order scrambling (2-col read across the gutter) — same words, wrong order = PASS-BUT-FALSELY at the corpus level. Validation gate = reading-order COHERENCE (no mid-sentence discontinuity), not just word-multiset. | PROPOSED |
 | **2 research sweeps owed** (not skills — project tasks) | note | The engine's signature capabilities are greenfield: (1) genealogy/temporal/citation-graph/KG-augmented/diachronic-NLP; (2) multi-voice persona-grounded attribution + voice-fidelity verification. Plus: validate claim-level NLI on multilingual/archaic humanities prose. | (project tasks, in research doc) |
+
+### Harvest 2026-06-28 (corpus settling — Musil graduated, Position I, gate fixed)
+
+| Element | Kind | One-line | Where it lands | Status |
+|---|---|---|---|---|
+| **`/graduate-chamber-source`** (already PROPOSED 06-26/27) | **EXPAND again** | Now carries the full EPUB path (`structure_from_ncx`→`insert_chapter_headings`→`clean_pandoc_html_residue`, used when `repair_epub_headings` coverage is low) + the **keep-notes** rule (convert footnotes/endnotes, never drop) + the per-book curatorial forks graduation always has (endnotes form, small-caps/two-line heading rendering, front/back trim, neighborhood). Empirical spec now = Musil (clean) + Taylor (the messy-EPUB case: two-line headings, raw `<a>`, `<sup>` footnotes). Build when the batch begins. | new skill | PROPOSED (expand) |
+| **The gate itself can be PASS-BUT-FALSELY** | verification-ladder entry | A verifier that checks an **enumerated set of cruft signatures** silently passes any residue outside the set — `verify_conversion` passed `cruft=0` on 59 raw `<sup>` (the cruft check only knew `audit_cruft`'s img/span-class/fenced-div patterns). When the gate gates a whole-corpus audit, its blind spot makes the PASS count untrustworthy. Antidote: gate on the **general class** (any residual raw HTML tag, tag-name-anchored) not just known signatures; periodically adversarially-probe the gate itself. Caught 2026-06-28; fix re-ran the audit 1187/91→1152/126. | `reference-verification-ladder.md` | **PROPOSED** |
+| **Symmetria §3 flag: finding-scoped-to-one-condition restated as unconditional verdict** | Symmetria §3 flag | A result true under a specific condition, restated as an unconditional rule, is contamination shape. Caught 2026-06-28: the method doc's "ocrmac more accurate than olmOCR" was scoped to **2-column figure-dense** (where olmOCR scrambles); I wrote "olmOCR RETIRED." Steward corrected (olmOCR is the single-column tool, more precise; complementary not ranked). Antidote: **verify the scope of a claim before generalizing it.** Kin to `trust-prior-pass-frame` (which is about test-scope; this is about claim-scope). | Symmetria §3 | **PROPOSED** |
+| **prose word-guard for faithful structure-cleaning** | verification-ladder entry | When a cleaner removes/reflows STRUCTURE (headings, printed titles, residue) but must preserve PROSE, gate it with a **prose-only word-multiset compare**: reduce both before/after to narrative words (exclude headings, numbers, printed-title runs, markup — anchored consistently on both sides), refuse to write on any non-zero delta. `clean_pandoc_html_residue`'s guard; proven raw↔final WORD-IDENTICAL on 464k-word Musil; caught every real defect across ~10 iterations. The complement to the byte-identical gate (which proves *output* identity; this proves *prose* identity under intended structural change). | `reference-verification-ladder.md` | **PROPOSED** |
+| **structure-from-authoritative-ToC = one engine, two map-producers** | method note (ladder or `/graduate-chamber-source`) | Answer to the long-open unify question: chapter-structure recovery is ONE placement engine (`insert_chapter_headings`: match anchor→unique line→replace-block/insert, fail-loud, dry-run) fed by TWO map-producers — the EPUB **NCX ToC** (`structure_from_ncx`, anchor=`<span id>`) and the OCR **printed page-number** method (ToC printed-page→scan-offset). Honour the work's *declared* structure; never re-infer. | `reference-verification-ladder.md` or the skill | **PROPOSED** |
+
+*(Five, all earned in use today. The gate-PASS-BUT-FALSELY and the prose-word-guard are the load-bearing two — the first is a defect found in the gold standard itself; the second is the reusable faithfulness instrument the whole Musil graduation rode. None created autonomously.)*