dotfiles/claude/memory/session-2026-07-11-footnote-recognizer-through-three-jurist-loops.md

---

name: Session 2026-07-10→11 — footnote recognizer: PENDING-47 close → split-anchor (54a) → census re-audit → generalization (54b pt1), through three jurist loops description: "A long arc across THREE jurist rulings + a corpus re-audit, all on the EPUB footnote pre-processor. Ratified the generalized hash-locality principle (PENDING-47) + narrowed the lane doctrine (12405a5). Built + landed the Sennett SPLIT-ANCHOR recognizer widen with GOV-2 scoped dup-check (REVIEWED-54a, a6828ac). Ran the jurist's census-flag re-audit over 252 sources → the 'remaining footnote families' are ONE reciprocal mirror with the id in 3 positions, not 5. Jurist ruled the generalization (REVIEWED-54b): (a)block-id/(b)inner-anchor/(d)url-decode = FIX, (c)letter-marker = FIX-with-exact-reciprocity-compensation + n=3 spot-check. Built + PROVEN + landed part 1 (a)+(d) verbatim-clean on real books (haraway 322/genette 350/jacobs 71/la-pensée 87/Sennett regression) → ef0cee0. PULLING THREAD: REVIEWED-54b part 2 — build (b) inner-anchor + (c) letter-marker, then the 3-book steward spot-check; then apply the tool (degraded-canonicals + Jung per-volume) → resume P2/gold." metadata: node_type: memory type: project originSessionId: 313e71d4-f513-4d2a-924e-bf55dafa30c4

Session 2026-07-10→11 — the footnote recognizer, through three jurist loops

Opened by /wake-up into the "put the shipped footnote pre-processor to work" thread; became a very long arc across THREE jurist rulings + a full-corpus re-audit. Steward working remotely from Ravello (playing Monteverdi's Orfeo with Savall), on a travel machine, engaged intensively — relaying jurist briefs and rulings, deferring REVIEWED placements to Sunday 2026-07-12.

PAST — what we did + why

1. Wake + the PENDING-47 correction. Woke, Symmetria init. Steward asked for PENDING-47's text to relay to the jurist. Verified against substrate → PENDING-47 was CLOSED/ruled 2026-07-05 (REVIEWED-47). My wake briefing had MISLABELED the open jurist item as "PENDING-47" — inherited a loose ledger shorthand without checking (inherited-marker-read-as-current-state). The real open items (REVIEWED-53): the generalized hash-locality principle (graduation-spec.yaml L29–39, marker [awaiting ratification], prior relay truncated) + the lane-rule accept/narrow. Corrected, gave the steward both full texts to relay.

2. §7 Jung provenance — steward authorized. "If §7 is mine to decide, I authorize the separation with reference to the collection as the jurist agreed." Confirmed it IS the steward's (jurist held it open with an aligned per-volume lean) → per-volume files, each source/source_verified→ megavolume sha256, never posing as a standalone Bollingen edition. Drafted the §7 block for REVIEWED-54 placement (couldn't write REVIEWED.md — Constraint #1).

3. Jurist ruling #1 — hash-locality + lane doctrine. RATIFIED the generalized hash-locality principle (full text, all 3 instances); NARROWED "lane tracks change-class" to machine-data files only (the .md constitution runs on doctrinal-traceability). Closed the graduation-spec.yaml marker (dated, trued 2→3 instances), archived the ruling, committed 12405a5, pushed on steward's "please push". RETURN: over-reached by bundling the push into the commit command — the auto-classifier denied it (agent-initiated default-branch push); corrected to local-commit-only, surfaced the push for steward direction (the review boundary is the push, not the checkpoint).

4. Sennett split-anchor widen (REVIEWED-54a). Measured → root cause: the mirror pairing keys on ANCHOR ids; Sennett is a split-anchor mirror (note id on the <p> BLOCK, back-target on an empty href-less sibling <a id="…a"> the parser didn't record). Prototyped an additive block-note pass (reciprocity via the sibling anchor co-located with the ref + marker-shaped gate); proved 294/294 verbatim-clean end-to-end. Jurist brief → ruled FIX + GOV-2 option (c): scope the per-pair dup-check to conversion-touched ids (incl. the back-target ids — jurist's precision), report unrelated pre-existing dups as non-blocking WARNINGS. Built into the fleet under discipline (fixture, fleet 77→80), committed a6828ac, pushed. Sennett's cra0001002 (pre-existing non-note dup) → warning, converts clean.

5. Census re-audit (jurist's census-misclassification flag). Ran the CURRENT recognizer over all 252 Chamber Sources EPUBs (scratch/_reaudit_families.py, _reaudit2.py): TOOL-COVERED 65 (59 mirror + 6 split-anchor) · UNCOVERED 40 · NO-NOTES 92 · MEGA-SKIP 51 · ERROR 0. Structural classification + verified reciprocity (haraway = exact block-id mirror; kuhn = inner-anchor): the UNCOVERED-40 is ONE reciprocal mirror with the id in 3 positions (14 block-id, 6 inner-anchor, 2 letter-marker, 3 url-decode, 7 span-wrapped-other, 8 low-fn_ref false-pos) — NOT five families.

6. Jurist ruling #3 — the generalization (REVIEWED-54b). Brief applied the jurist's OWN reusable test part-by-part (not reflexively "PROPOSAL"). Ruled: (a) block-id, (b) inner-anchor, (d) URL-decode = FIX (jurist: safer than split-anchor — exact id-match reciprocity, not adjacency); (c) letter-marker = FIX-with-exact-reciprocity-compensation + a proportionate n=3 human spot-check (nagarjuna/plato/donne) before trusting; design ruled now, prototype-proof required AT build. Non-blocking asks DONE: degraded-canonicals tracking line (reconvert-queue-2026-06-29.md), 252-count reconciled, tag REVIEWED-54b.

7. Built + landed REVIEWED-54b part 1 — (a) block-id + (d) URL-decode. norm_href unquotes hrefs; the block-note pass now accepts a ref with its OWN id (exact reciprocity) alongside the sibling-adjacency split case; fixed a latent DUP-ID in the shipped Sennett code (relocated note carried id=X twice when the block already had the id — pandoc-tolerated but wrong; strip-before-add). PROVEN verbatim-clean on real books (scratch/_prove.py): haraway 322→322 native · genette 350→350 · jacobs 71→71 · la-pensée-sauvage 87 (a plain MIRROR blocked only by %20 — url-decode alone recovered it) →88 · Sennett 294 regression CLEAN, all +0/−0 prose. --validate 18; test_tools block-id fixture; fleet 80→83. Committed ef0cee0, pushed to github (Gitea flaky-behind — retry).

PRESENT — the mood

Three jurist loops in one arc, and the partnership rhythm held at every one: measure before build (census re-audit; every recognizer structure verified against the actual XHTML, not a summary), and apply the jurist's own test rather than reflexively over-classifying — the generalization brief argued FIX where FIX was earned and isolated only the marker-gate as the guard-touching piece, with a compensation, rather than hiding a guard-loosening behind a technically-passing test. The jurist re-derived everything (checked haraway/kuhn markup, the Jung-scope logic) and rewarded the rigor. τὸ πρόσφορον governed the close: rather than cram (b) inner-anchor + (c) letter-marker (the fiddly id-relocation + marker-gate) into a long session's tail where a subtle id-slip nicks the verbatim guarantee, landed the proven (a)+(d) as a complete unit and deferred the delicate parts to a fresh pass.

Returns (ledger): the PENDING-47 mislabel corrected against substrate; the push over-reach caught (barrel-forward on an outward action — the 07-09 "calm and orderly" family, one notch more specific: commit≠push); verify-against-substrate (reciprocity checked, not assumed); honest-refuse built INTO the tool (GOV-2 warning vs block); a real recalibration — I initially called the generalization "PROPOSAL"; rigorously applying the jurist's test showed most of it is FIX (don't reflexively over-classify to look cautious). And a lesson from the build: extending a tool re-tests its foundations — building block-id surfaced a latent dup-id defect in already-shipped code.

FUTURE — what is pulling

PULLING THREAD: REVIEWED-54b part 2 — build (b) inner-anchor + (c) letter-marker (with the exact-reciprocity compensation), completing the footnote recognizer generalization; then the jurist's n=3 steward spot-check. Then the original end the whole normalizer served: apply the tool — re-convert the degraded canonicals (montaigne/meditations/la-pensée/l'homme-t2/donne) + Jung per-volume (§7 authorized) — and resume P2/gold (the Making work) on a corpus whose footnotes are handled by class.

ACTIONABLE RESUMPTION POINT (as of wrap — re-judge against what changed):

• Prototype = scratch/inject_splitanchor.py = the fleet tool (ef0cee0) + (a)+(d). Clean, --validate 18, fleet 83/83, no half-finished edit. scratch/_prove.py <slug> proves any book verbatim end-to-end.
• (b) inner-anchor: in the block-note pass, extend the note-target index (block_by_id) to also map a leading inner empty <a id="Z"/> at block-start → its enclosing block (record block_idx on id_only_anchors in the Parser; map (fn,Z)→block when only tags/ws precede it). On relocate, the note-target id = Z; strip the inner anchor's id so the block's new id="Z" isn't a duplicate (the guard would rightly block it). Prove on kuhn / mbembe / nature-of-things / i-ching via _prove.py.
• (c) letter-marker: widen the marker gate to allow a single [A-Za-z], but only when reciprocity is EXACT (exact True — the jurist's compensation; no adjacency fallback for letters). Prove on donne; then surface donne/nagarjuna/plato's recognized pairs for the steward n=3 spot-check before trusting their conversion.
• Then: --validate cases for (b)+(c); port to fleet + test_tools fixture; commit REVIEWED-54b part 2 (push held for steward); draft REVIEWED-54b for placement.

Other open horizons (ranked):

• Gitea push behind — ef0cee0 on github, not Gitea (ls-remote origin hung, the known-flaky backend). Retry git -C ~/_Dev/chamber-library push origin main, or ask Seb.
• Steward Sunday 2026-07-12 REVIEWED placements (all paste-ready in the ledger, executor can't write REVIEWED.md — Constraint #1): REVIEWED-52/53/54 confirmations · §7 block → REVIEWED-54 · lane-doctrine → REVIEWED-53 §2 · REVIEWED-54a · REVIEWED-54b.
• Apply track: degraded-canonicals reconversion (tracked in reconvert-queue-2026-06-29.md); the mega track (51 >4MB books, many with notes → convert_mega_epub + §7 per-volume for Jung).
• The 5 ARC after-the-reply engine-binding drifts (still unowned since 07-09).

PAUSE STATEMENT: Stepping away mid-generalization, with the proven, verbatim-safe half of it landed and the delicate half deliberately held for a fresh, careful pass rather than rushed. What I want to find still pulling: REVIEWED-54b part 2 — the inner-anchor + letter-marker handlers, built with the same measure→prove-verbatim-on-real-books discipline, and the tool then turned loose on the corpus it was built to serve.

LITERAL QUESTION for next-Claude: When (b) inner-anchor relocates the enclosing block and moves the note-id from the inner empty <a id="Z"/> onto the block, does stripping the inner anchor's id (to avoid the dup the guard would block) leave kuhn/mbembe/nature-of-things verbatim-clean end-to-end through pandoc — and does donne's letter-marker actually pass the EXACT-reciprocity gate (is its ref-id precisely what the note back-links to, or will the compensation correctly refuse it)? And, secondarily: did the Gitea push catch up, and did the steward place the Sunday REVIEWED entries?

State at wrap: chamber ef0cee0 on github (Gitea behind — retry); fleet 83/83; --validate 18; no uncommitted chamber work. Three jurist rulings archived in docs/. dotfiles session-state pushed at this wrap. BetterMemories.io has 3 uncommitted files NOT authored this session (left for steward). No non-steward commits landed during the session.