session 2026-07-09: Handke de-gold graduation + archive_sources --add mode + PENDING-52; V2-review session memory + drift-patterns + skill-harvest

claude/memory/skill-harvest-register.md

@@ -381,3 +381,9 @@ The single place proposed skills live so they don't evaporate between sessions.
 | **cross-volume verify-before-delete move** | verification-ladder note | Moving data across filesystems (internal→external cold archive): `rsync -a` → verify **exact file-count match + `du`** (NOT a byte-sum — mine computed 0=0 vacuously) + rsync rc=0 → only then `rm` the source. Isolate flaky-remote pushes so a hang can't block. And **check git's real exit, not a pipe's** (`… | tail` masks push failure — bit me on the dotfiles→gitea push this wrap). Proven twice today (MemPalace cold-archive). | `reference-verification-ladder.md` | **PROPOSED** |
 
 *(Both genuine — /model-handoff is load-bearing: it just proved itself and will again next session. The byte-sum-vacuous + pipe-masks-exit catches are also logged as KG drift-patterns; not manufactured.)*
+
+### Harvest 2026-07-09 (V2 review + Handke de-gold graduation + the read-repo-docs-first lesson)
+- **PROPOSE — Symmetria §3 flag: `enter-a-repo → read its CLAUDE.md + the runbook/spec it names, BEFORE acting`.** Recurrence met HARD: I started graduating in chamber-library WITHOUT reading its local CLAUDE.md first (a standing global directive), and re-derived — badly, with wrong source-location guesses — a fully-documented process (attest-never-default · source-archiving-IS-a-graduation-step · conversion-runbook · graduation-spec). The steward corrected ~4× in a row, each the same shape ("this is in the runbook / the CLAUDE.md"). **Flag text:** *before acting in a repo, read its local CLAUDE.md and the runbook/spec it points to — re-deriving a documented process is the contamination shape (looks like work, produces drift, spends the steward's corrections).* Kin to `resurface-banked-notes-before-rederiving`. **PROPOSED — recurrence-promoted, not minted-from-one.**
+- **PROPOSE — Symmetria §3 flag / feedback: `provenance-is-part-of-the-process, not a loose end`.** I framed source-attestation + archiving as "loose ends / deferrable"; the steward: *"they should never be considered loose ends, they are part of the process."* Executor bias: treat the visible artifact as done, provenance (attestation, archiving, cataloguing) as afterthought. **Text:** *attestation, source-archiving, cataloguing are constitutive steps, not the tail; the feeling that "the artifact is done" hides the provenance the next hand needs.* Kin to `completion-is-a-tripwire`. **PROPOSED.**
+- **`/graduate-chamber-source` — REINFORCED (the born-digital/EPUB branch, full pipeline incl. archive+attest+catalogue).** The Handke run is the born-digital branch (distinct from Levi-OCR + Olson-V-SCAN): identity-verify → pandoc + structure pass → gates → archive via `archive_sources.py --add` + steward `source_verified` + catalogue. The skill, when built, must carry the branch-selection heuristic AND make archive+attest+catalogue non-skippable (the whole "loose end" failure this session). **PROPOSED (reinforced; build-on-need — another strong nudge it's ripe).**
+- **PROPOSE — verification-ladder entry: `a red gate deserves the same scrutiny as a green one`.** verify_conversion's `heading_density` FAILED the Handke (a genuinely continuous work). Before trusting OR overriding, I read the gate's own §VII sectionless exception + how the English sibling passes → the honest fix (title-as-`#` heading matching the sibling; NOT invented chapters, NOT a bare `sectionless:true` flag). **Rule:** a gate FAILURE is a candidate, not a verdict — verify true-defect vs heuristic-false-positive (read the gate's own escape conditions + a passing sibling) before either trusting or overriding it. The green-gate-blindness lesson's mirror. Home: `reference-verification-ladder.md`. **PROPOSED.**