session 2026-07-09 evening: PENDING-53 (cross-repo re-anchor gap) + P1-clean/EPUB-normalizer-pivot session memory + KG drift-patterns

claude/memory/session-ledger-2026-07-09.md

@@ -14,6 +14,8 @@ metadata:
 
 - 10:35 — Fable's V2 design doc is REAL (thread now **confirmed**, not superseded). Did the jurist-review same-seat + independent substrate re-verification rather than trusting the doc's "every claim verified" self-attestation (holding the day's two drift-patterns). Result: §1 substrate findings reproduce **exactly** (1124 escaped apostrophes; Mauss 943/864/817; German-source finding; G&G Thibon paratext). Sole correction: §1.5 manifest census says 13, substrate = 21 (→ matches plan; P4; immaterial). Fence check (V0 §5) passed: uniform-20% first-batch distribution is compliant. Converges with the actual jurist's relayed review (don't redo with Fable).
 
+- 17:16 — **Wake (post-clear) into the same-day thread.** Thread-validity = **confirmed** (V2 gold prep; P3 done, P1 next). Caught the day's own `swallowed-tool-error-as-absence` drift **twice in the wake itself**: (1) `ls session-*.md` returned empty → did not read as "no session files"; (2) `ls session-ledger-*.md` returned empty → **checked `test -f` before Write**, found today's ledger EXISTS — overwriting would have destroyed the prior entries. The drift-pattern I woke holding is the one I nearly repeated; naming it here so the return strengthens, not just avoids. Holding for the session: read-repo-docs-first + calm/plan-first before P1.
+
 ## Authorization moves
 - 10:40 — Folded 2 jurist-requested mechanical fixes into the V2 design doc (D-1, recorded not silent): §2.2 cross-ref §11.4→§11.3; added precondition P7 (stratum-tag pre-existing gold). Owed REVIEWED entries: (a) charter's V0-gate + œ/æ ratification (still unfiled, steward); (b) this V2-design jurist-review outcome.
 
@@ -34,3 +36,16 @@ metadata:
 ## Sub-agent dialogues
 
 ## Bypasses
+
+## P1 — L'Enracinement source-clean (post-clear, read-repo-docs-first held)
+- 17:55 — **Read the governing docs FIRST** (chamber CLAUDE.md + conversion-runbook + graduation-spec + V2 §14) — the exact step 2026-07-09-daytime skipped. Then answered the wake's literal question with substrate evidence: re-hashing `lenracinement.md` breaks **no hand-bound anchors** — only `catalogue.yaml:11471` carries the sha (regenerable), no reading-index (only 4 exist, none Weil), no phase-2 citation. Re-anchor = one `build_catalogue.py`.
+- 17:56 — **Finding (would have shipped past on a naive strip):** the C1 defect is TWO classes, not one — 6,671 `\'` + 26 `\.` (pure pandoc escapes, faithfully removable) **plus `\[1\]`×2 = Weil's footnote 1** (marker L1279 + def L2624), which a blanket `\[→[` would half-break. keep_notes governs → convert to `[^1]`/`[^1]:`. The measure-whole-first discipline caught it (the frontmatter `title:` also carried `\'`).
+- 17:57 — **Green-gate insufficiency compensated:** `prose_delta` is a word-multiset → structurally blind to the punctuation edit → reads 0 regardless. Did NOT rest on it; added a char-level faithfulness alignment (0 non-backslash divergences) + pandoc footnote-pairing render check. (Holds today's `green-gate-reads-as-thoroughness` + `verification-passed-vacuously` patterns.)
+- 18:00 — **APPLIED (steward-authorized both recommendations: scoped in-place + footnote-in-P1).** Bounded-change proof: prose_delta 0; char delta 6698; residual backslashes 305; 1125 lines changed; verify_conversion **5/5 PASS**; catalogue re-anchor bounded to 2 lines (title+sha), drift gate in-sync; pandoc fn1 pairs. New sha `58ead0ab…` (was `0553bd5f…`). Footprint: canonical .md + catalogue.yaml, both **uncommitted** (governed repo + Gitea flaky → steward's commit call, joining the prior 4). Backup at `scratch/making/lenracinement.pre-p1.bak`.
+- **NEXT (do not barrel):** P2 (G&G sidecar + EPUB-residue clean), then P4/P5/P6/P7 — steward directs the rhythm.
+
+## Return — P1 re-anchor was INCOMPLETE (caught during P2 diagnosis)
+- 18:20 — **Named + fixed my own miss.** In P1 I claimed "only catalogue.yaml binds to the sha" — but I scoped the grep to **chamber-library only**. The **engine** carried 4 more bindings to the old sha (`corpus/manifest.yaml:184`, `coverage-ledger.json:316/330`, `sidecars/weil-enracinement.meta.json:5`). The wake's literal question said "grep what is bound to lenracinement's sha" — the binding surface spans **both repos**; I checked one. Textbook `trust-prior-pass-frame` / verify-at-the-scope-of-the-extension. Found it only because P2 diagnosis opened the engine sidecars — luck, not rigor. **New standing check: after any chamber canonical re-hash, grep BOTH chamber AND studium-engine for the old sha.**
+- 18:22 — **Fixed correctly by mechanism:** manifest.yaml + sidecar are curated → hand-fixed the declared sha (assert 1 occ each); coverage-ledger is `generated_by: ingest_gate.py` → regenerated via the gate (recomputes live sha; §1.1 hash-binding is the proof). Result: 0 old-sha remain; weil-enracinement **VALIDATED**; line count unchanged (2624) so sections/`source_lines` stayed valid.
+- 18:23 — **Pre-existing finding surfaced (NOT mine, NOT fixed):** regenerating the ledger revealed **5 latent failures — `after-the-reply-i…v`** (ARC essay reading-sources; manifest+sidecar sha drift from ARC revisions). The stale ledger had hidden them behind a false "validated: 15". Honest-degradation win, but a separate ARC↔engine re-anchor task to flag. Engine now: validated 8 / failed 5 (the 5 ARC) / blocked 0.
+- Engine footprint (uncommitted, D-1): manifest.yaml (1 line), weil-enracinement.meta.json (1 line), coverage-ledger.json (regenerated).