obsidian-plugin/lllin000_PaperForge
1
0
Fork 0
mirror of https://github.com/lllin000/PaperForge.git synced 2026-07-22 06:50:53 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions
lllin000_PaperForge/PROJECT-MANAGEMENT.md

42 KiB
Raw Blame History

OCR-v2 Project Management Log

Branch: master | Last Updated: 2026-07-02 Active work: Figure caption prefix recovery + inline HTML fix + table matching verification completed. 585 regression tests pass.

0. Executive Summary

Current state: All 10-paper audit issues resolved (3 audit fix commits + rotated prematch refactor + asset-internal figure number recovery). U746UJ7G figure_unknown_000 → figure_002 with recovered label "Plot of Criteria Time". 428 targeted tests pass. Next: Monitor production OCR for regressions, archive stale project docs.

1. Architecture

1.1 The problem (pre-v2)

raw_label/text → final role → normalize → rescue → demote/promote → attach → render

Early role guesses caused: title/author/footnote misclassification, body paragraphs typed as references, reference continuations leaking into body, left/right column interleave bugs, duplicate figure captions.

1.2 The solution (v2 target)

raw observations → structural signatures → stable anchors/families → zone inference → role resolution → figure/table validation → render + health

1.3 Core principles

  1. seed_role is a proposal, never a final role — assign_block_role() proposes; normalize_document_structure() decides
  2. VERIFY_REQUIRED roles (paper_title, authors, section_heading, reference_item, figure_caption, etc.) must have role_verification_status == "ACCEPT" with non-empty role_source and role_evidence
  3. Zone != Role — being in the body reading environment does not imply body_paragraph
  4. Reference tail first — reference sections protected from tail contamination
  5. Frontmatter source-backed — OCR localizes but must not invent/discard canonical frontmatter

Key design documents:

  • Architecture spec: docs/superpowers/specs/2026-06-08-ocr-anchor-first-structured-parsing-design.md
  • Implementation plan: docs/superpowers/plans/2026-06-08-ocr-anchor-first-structured-parsing-plan.md
  • Role gate plan: docs/superpowers/plans/2026-06-11-ocr-verified-structural-role-gate.md
  • Latest spec realignment: docs/superpowers/specs/2026-06-13-ocr-real-paper-regression-and-spec-realignment-design.md

2. Current Status

2.1 Test Suite


| Suite | Result | |-------|--------| | Figure stack (figures + reader + containment + backmatter boundary) | **286 passed** ✅ | | Document + roles + render + gate + rebuild + state machine + trace | **330 passed, 0 failed** ✅ | | Author bio detection (ocr_bio) | **37 passed** ✅ | | Spec contracts | All passed ✅ | | Real-paper regressions | All passed ✅ | | **Total OCR tests** | **1018 passed, 275 skipped (fixture unavailable), 0 failed** ✅ |
| Structural gate | Installed | | Role assignment (seed only) | ✅ | | Zone inference + fallback | ✅ | | Figure inventory | Global distance clustering — correct | | Backmatter boundary | ✅ Ref-anchored partition committed | | Pre-ref tail zone | ✅ Fixed | | Abstract detection | ✅ Fixed | | Frontmatter heading normalization | ✅ Fixed — no text matching | | non_body_insert clustering | ✅ Fixed — page-1 guard threshold | | Render (backmatter headings) | ✅ Correct heading rendering | | State machine | ✅ Accepts done_degraded as terminal | | Rebuild (span backfill skip) | ✅ Version match check fixed | | Author bio detection (Passes B+C) | ✅ P0: post-ref text-only bios as backmatter_body. P1: figure-residual portrait assets as author_bio_asset + figure_caption support |

2.3 Fix Status

# Paper Issue Type Fix Commit
1 — P21 zone fix (4AG67PBH "Conflict of Interest" heading) Pipeline code infer_zones() frontmatter_side_blocks page gate: added first_reference_page is not None and page >= first_reference_page - 1 35aabae
2 4AG67PBH Author bio text in post-ref reference_item (b8/b10) New module post_ref_bio_cleanup reclassifies bios as backmatter_body; _bio_text_score category-weighted 0-5 scoring e2f0c8a
3 — author_bio_asset role contract Pipeline code author_bio_asset added to render_default=False, index_default=False skip sets 7810eb1
4 — Pass C pipeline wiring Pipeline code Insert post_ref_bio_cleanup + prune_figure_inventory_after_bio after write_back_figure_roles 7810eb1
5 — P1 residual author bio pass (Pass B) New function residual_author_bio_pass detects portrait unmatched_assets/unresolved_clusters with nearby bio text 7a1cc5e
6 — P1 figure_caption support in Pass C Role expansion post_ref_bio_cleanup extended for figure_caption role 7a1cc5e
7 — tag_figure_contained_text author_bio guard Protection Skip author_bio blocks and author_bio_asset role in figure containment 7a1cc5e
8 — Bio word limit 80→200 Pipeline code Real bio text 90-100 words exceeded 80-word limit in _bio_text_score ae081a4
9 4AG67PBH Barbara bio as structured_insert_candidate missed by Pass C Role expansion Added structured_insert_candidate to Pass C role list ae081a4
10 4AG67PBH Page 25 portrait id=5 missing from unmatched_assets Pipeline bug fix ocr_figures.py 4479/4529 used page-agnostic bare block_id filter, hitting collision when same id exists on different pages. Changed to (page, block_id) tuples. ae081a4
11 4AG67PBH Acknowledgment text absorbed as reference_item (p21) Regex fix _is_reference_item_candidate tightened fe9cc70
13 WV2FF4NV Fig 10/6 locator caption bridge New feature _is_previous_page_legend_locator + bridge in build_figure_inventory: connects locator → previous full legend → current visual group. Recovers misclassified legends from rejected_legends. 3f61f4a
14 — Issue 3: Backfill word leakage beyond bbox Word-level filter Added _word_belongs_to_block + _word_center_inside_rect filters after expanded clip 796e8bb
15 — Issue 1A: Validation-first bare table skips same-page asset Continue guard Validation-first branch only early-exits when no same-page assets exist 796e8bb
16 — Issue 1B: Split table caption continuation escapes ownership Continuation materialization _find_table_caption_continuation + _materialize_table_caption inside build_table_inventory 796e8bb
17 — Issue 4: Short papers (≤2p) incorrectly red + needs_rebuild Health profile Added _health_profile(page_count) → short_form waives abstract/heading gates 796e8bb
18 — Issue 2: Demoted-caption figure inner-text leakage Container bbox regions Validated _container_bbox regions in tag_figure_contained_text via 3 helpers + containment-only integration 0e4ecbc
19 — Issue 5: Cross-column page_assets groups falsely accepted Column-homogeneity gate _column_band_id + rejection in _is_safe_page_assets_group 4ab227e
20 — Issue 6: Same asset consumed by figure AND table Post-hoc arbitration resolve_media_asset_conflicts resolves asymmetric cases; weak/weak stays in ownership_conflicts 4ab227e

P2#1a — Previous-page legend locator bridge (✅ Fixed 3f61f4a)

WV2FF4NV Fig 10: locator "See legend on previous page" on p16 not bridged to full legend (in rejected_legends, misclassified as body_paragraph)

Remaining P2

  1. Figure containment gaps — cluster_bbox containment only runs on matched figures; composite figures with demoted caption upstream never enter containment. Inner-text detection misses vision_footnote/paragraph_title raw_labels.
  2. Short "Table N" caption matching — bare "Table 1." labels miss table_asset in ownership pipeline. Existing code partially handles it; residual cases remain.
  3. Page-1 body paragraph OCR text extraction failure (was: width check) — W33NLVJ2 right col Introduction block: OCR detected text region (raw_label=text) but extracted empty string. Pipeline previously mis-signaled as unknown_structural. Fixed: now ocr_text_missing.
  4. Body text backfill overlap — backfill_missing_text_from_pdf uses get_text("words", clip=expanded) returning words beyond block's bbox; can cause text duplication at render level.
  5. 2HEUD5P9 reference ordering — ✅ Fixed by F12 (raw_label=reference_content signal)
  6. Mixed-column tail page zone absorption — ✅ Fixed: tighten year-period regex

P3 (Boundary / Edge Cases)

  1. DW biography page mismatch — pages 32-34 vs expectations 33-34.
  2. AJR side-caption recovery — deliberately deferred; not part of generic inventory refactor.
  3. Chinese Windows encoding — non-ASCII PDF filenames garbled via GBK codepage. Glob fallback added; residual meta.json corruption from earlier runs.
  4. backfill_missing_text_from_pdf bbox-exact filtering — need decision on render-level dedup vs backfill-level bbox clamp.
  5. Multi-caption page column collapse — page_assets group on multi-column pages merges separate figures in different columns.
  6. Figure/table shared-consumption registry — no shared consumed registry for ambiguous image-like blocks.
  7. Short papers (<3 pages) health falsely red — no headings/abstract in Letter/Editor formats.

4. Active Queue

  1. ✅ P1 backmatter boundary (ref-anchored partition)
  2. ✅ Pre-ref tail zone fix (4KCHGV2Z)
  3. ✅ Gate 5 frontmatter fix series (24YKLTHQ)
  4. ✅ Stale trace-vs-expectation fixtures cleared (10 assertion updates)
  5. ✅ All stale test expectations reconciled (non_body_insert, caffard, legend_like, structural gate, render, state machine, rebuild, truth docs)
  6. ✅ P0 author bio detection (post_ref_bio_cleanup for reference_item)
  7. ✅ P1 author bio detection (residual_author_bio_pass + figure_caption support + tag_figure_contained_text protection)
  8. NEXT: Run lint (ruff) then merge ocr-v2 → master

4.1 Immediate Next Steps

  • 25 fixed tests (10 distinct issues)
  • Full OCR regression sweep: 1018 passed, 0 failed
  • P0 bio detection: 30 new tests, 0 regressions
  • P1 bio detection: 7 new tests, 0 regressions
  • tag_figure_contained_text author_bio protection
  • Run lint (ruff)
  • Merge ocr-v2 into master

5. Key File Map

5.1 Production Code (OCR Pipeline)

File Role
paperforge/worker/ocr_blocks.py Raw + structured block generation; preserves seed_role only
paperforge/worker/ocr_roles.py assign_block_role() — seed/proposal logic only (NOT final)
paperforge/worker/ocr_document.py normalize_document_structure() — anchor/family/zone + gate + final roles
paperforge/worker/ocr_structural_gate.py VERIFY_REQUIRED role decision + abstract span + reference zone + health
paperforge/worker/ocr_orchestrator.py Body reorder, column validation, layered assembly
paperforge/worker/ocr_render.py render_fulltext_markdown() — consumes verified artifacts ONLY
paperforge/worker/ocr_health.py Health reporting, merged gate summary
paperforge/worker/ocr_profiles.py Span extraction, profile aggregation, cross-validation
paperforge/worker/ocr_figures.py Figure reader pipeline
paperforge/worker/ocr_tables.py Table inventory matching
paperforge/worker/ocr_scores.py Score functions (spatial, structured_insert, etc.)
paperforge/worker/ocr_rebuild.py Derived rebuild entry point
paperforge/worker/ocr_pdf_spans.py PDF span backfill for OCR-missed blocks
paperforge/worker/ocr_bio.py Author biography detection utilities and passes

5.2 Test Files

File Purpose
tests/test_ocr_figures.py Figure inventory, matching, ownership
tests/test_ocr_document.py Document structure, normalize, gate
tests/test_ocr_render.py Fulltext render contract
tests/test_ocr_figure_reader.py Reader contract
tests/test_ocr_trace_vs_expectations.py Real-paper trace vs expectations gap report (8 gold papers)
tests/test_ocr_real_paper_regressions.py Page-level + document-level regression on real papers
tests/test_ocr_real_paper_audit_contracts.py Gold-fixture quality gate
tests/test_ocr_spec_contracts.py Architecture contract tests
tests/test_ocr_structural_gate.py Gate unit tests
tests/test_ocr_v2_structural_regressions.py Structural regression guards
tests/test_ocr_layout_first_regressions.py Layout-first behavior guards

5.3 Test Fixtures

tests/fixtures/ocr_real_papers/{CAQNW9Q2,DWQQK2YB,TSCKAVIS,A8E7SRVS,K7R8PEKW,6FGDBFQN,SAN9AYVR,2GN9LMCW}/

5.4 Design Documents

File Purpose
docs/superpowers/specs/2026-06-08-ocr-anchor-first-structured-parsing-design.md Architecture spec
docs/superpowers/specs/2026-06-13-ocr-real-paper-regression-and-spec-realignment-design.md Current phase spec
docs/superpowers/specs/2026-06-10-ocr-figure-reader-contract-design.md Figure reader spec
docs/superpowers/specs/2026-06-23-ocr-visual-grammar-hardening-design.md Visual grammar hardening
docs/superpowers/specs/2026-06-27-figure-containment-and-backmatter-boundary-design.md Current P0/P1 spec
docs/superpowers/plans/2026-06-18-ocr-v2-readiness-master-plan.md Readiness gates master plan
docs/superpowers/plans/2026-06-17-ocr-v2-closeout-single-plan.md Close-out single plan
docs/superpowers/plans/2026-06-15-group-first-figure-inventory-plan.md Group-first figure refactor (deferred)
docs/superpowers/specs/README-ocr.md OCR design index

5.5 Active Truth Files

File Role
project/current/ocr-v2-active-queue.md Active queue — next-work authorit
project/current/ocr_rebuild_audit.md Evidence source for queue
project/current/ocr-v2-generalization-boundary.md Architecture boundary note
project/current/ocr-v2-remaining-issues-2026-06-18.md Historical readiness residuals

6. Decision Log

Date Decision Rationale
2026-06-08 Adopt anchor-first OCR architecture Early role guessing caused cascading errors; needed structural discovery first
2026-06-10 Separate figure reader from body prose Figure info must be reader-visible without body pollution
2026-06-11 Install verified structural role gate Spec was being bypassed; gate enforces seed≠final contract
2026-06-13 Dual-gate regression + spec-contract testing Tests protected helper behavior, not real-paper outcomes
2026-06-13 Root-cause approach: no renderer patches Heading merge goes in raw blocks; boundary detection fixed in infer_zones
2026-06-13 Figure sequential matching as cross-page tradeoff Caption-asset pairs on different pages get lower confidence
2026-06-15 Expand deterministic gold set to 8 papers Needed broader regression surface before changing figure inventory
2026-06-15 Do not solve AJR side-caption recovery in group-first refactor Keep scope generic; AJR-specific rescue is later phase
2026-06-15 Group-first matching is next architectural target Existing clusters/visual groups too late in pipeline
2026-06-17 Single-thread close-out note authoritative queue moved to project/current/ocr-v2-closeout-priority.md
2026-06-21 Replace greedy region-growth with global distance clustering Human sees assets as perceptual groups, not competing candidates
2026-06-23 Pre-ref=body flow, post-ref=backmatter CRediT/Ethics above References → body_zone
2026-06-23 Figure containment: render-hygiene pass build_figure_inventory Containment shouldn't affect matching, only rendering
2026-06-23 Reference sort: two capture groups Prevents false matches on plain year numbers
2026-06-26 P0 before P1 Fix 2/4/5 are <30 lines fully understood; Fix 1/3 need new specs
2026-06-28 Pre-ref tail zone: strip from region_bus not re-apply _apply_zone_labels re-applied stale tail zone from infer_zones() after ref partition. Fix: strip pre_ref block IDs from region_bus before zone re-apply.
2026-06-28 Author byline: require lowercase letters _looks_like_initial_lastname_byline matched all-caps journal taglines. Fix: require any lowercase letter in matched text.
2026-06-28 Page-1 body_start: metadata headings should not trigger _is_first_page_body_start treated ANY section_heading as body start. Fix: only real body section headings (introduction, methods, etc.) trigger body_start on page 1.
2026-06-28 Frontmatter heading normalization: no text matching Metadata sidebar labels rejected by structural gate fell to unknown_structural. Fix: normalize held heading blocks in frontmatter_main_zone to frontmatter_noise using only zone + gate decision + seed_role, no text matching.
2026-06-28 Author bio detection: three-pass cascade, P0 first Strong structure first, residual explanation second. Real figures must never be preempted. P0: post-ref text-only. P1: figure residual. P2+: P1 profile card pre-pass.
2026-06-28 Category-weighted bio scoring career=+3, education=+2, research=+2, institution=+1, publication=+1. Returns (score, categories) tuple. Threshold: score ≥ 4 AND categories ≥ 2.
2026-06-28 author_bio_asset role: non-rendered, non-indexed Bio artifacts removed from figure_inventory entirely, never returned to unmatched_assets. Clean prune before reader.
2026-07-01 Asset-internal figure number recovery: metadata-only pass Recovery must NOT split OCR blocks or mutate chart text — only patches figure_number, figure_id, recovered_label_text on existing matched figures. Coordinate normalization is caller responsibility.
2026-07-01 Broaden recovery gate to handle normal prematch unknown figures Synthetic-figure gate (bbox_only_asset flag) excluded figure_unknown_NNN from normal rotated prematch path. Gate now allows figure_unknown figures without synthetic flags.

7. Agent Instructions

7.1 Project Folder Management

The authoritative prompt for project record management lives in .omp/AGENTS.md. It is auto-loaded by omp into every session context and defines when and how to update PROJECT-MANAGEMENT.md, project/current/*, and project/archive/.

TL;DR: PROJECT-MANAGEMENT.md updated every session end. project/current/ updated at milestones only. project/archive/ gets moved-to (not deleted) when stale.

7.2 How to Update PROJECT-MANAGEMENT.md

  1. Update section 2 (Current Status) — test counts, component status, fix status
  2. Update section 3 (Remaining Issues) — remove resolved items, add new ones
  3. Update section 4 (Active Queue) — check/adjust next steps
  4. Add new decisions to section 6 (Decision Log)
  5. Add a compressed entry to section 8 (Session Timeline)
  6. Update "Last Updated" date at top

7.3 Before Starting Any OCR-v2 Work

  1. Read section 4 (Active Queue) for current priority
  2. Read the relevant design doc from section 5.4
  3. Understand what the tests currently expect
  4. Work one repair at a time; verify before moving on

7.4 Test Commands

# Full OCR test suite
python -m pytest tests/test_ocr_*.py -v --tb=short

# Figure stack only
python -m pytest tests/test_ocr_figures.py tests/test_ocr_figure_reader.py tests/test_ocr_render.py -q
# Real-paper regression only
python -m pytest tests/test_ocr_real_paper_regressions.py -v --tb=short

# Spec-contract tests only
python -m pytest tests/test_ocr_spec_contracts.py -v --tb=short

# Structural gate only
python -m pytest tests/test_ocr_structural_gate.py tests/test_ocr_document.py -v --tb=short

# Rebuild a single paper
python scripts/dev/ocr_rebuild_paper.py DWQQK2YB

# Rebuild + regenerate block_trace
python scripts/dev/ocr_rebuild_paper.py --trace DWQQK2YB
# Lint
python -m ruff check paperforge/worker/ocr_*.py

7.5 Design Doc Reading Order

  1. docs/superpowers/specs/README-ocr.md — index
  2. docs/superpowers/specs/2026-06-08-ocr-anchor-first-structured-parsing-design.md — architecture
  3. docs/superpowers/specs/2026-06-13-ocr-real-paper-regression-and-spec-realignment-design.md — current phase

8. Session Timeline (Compressed)

Date Session Key Results Detailed Archive
2026-06-16 Gold fixture expansion + 10 pipeline fixes 98 bug annotations, 8 papers audited, 8 fixes applied (F1-F10) §9.1
2026-06-17 OCR-v2 boundary close-out pass Zone-boundary fixes, tail/backmatter shrink, correspondence routing. 202P/1F/43S §9.2
2026-06-18 Readiness-gates implementation Gates 1-4 complete + blind audit protocol. 249/249 figure/health/document pass §9.3
2026-06-19 Blind audit (5 unseen papers) All PASS — no new failure families. OCR-v2 declared "state healthy" §9.3
2026-06-21~22 Figure merge: greedy → global distance clustering Union-find clustering, 261 tests. Caption-independent semantic grouping, ownership registry, local pairing, conflict detection §9.5
2026-06-22 Cross-page caption consumption fix Reader/render contract breach fixed — cross-page matches now consume caption on legend page §9.4
2026-06-23 Visual grammar hardening Composite parent detection, dense page arbitration, figure/table separator veto, dedup refinement §9.5
2026-06-26 Rebuild production run (699 papers) --resume checkpoint, glob fallback, N6XCZD25 body text fix §9.6
2026-06-26 Figure number inference + container admission Leading [1] gap filled, blue sidebar box rendered as [!NOTE] §9.7
2026-06-26 UI polish (plugin) Dashboard CSS, maintenance tab redesign, Vercel-style polish §9.8
2026-06-27 Rebuild audit + index repair 6 hard-block/high bugs fixed (sync, workspace, field registry) §9.9
2026-06-27 Deep investigation — 5-fix spec P0 all committed (ref sort, caption insert, figure containment). P1 backmatter in progress §9.10
2026-06-28 P1 backmatter boundary committed Ref-anchored partition (3e33e5b). Pre-ref=body flow confirmed (9b72783). 16/16 tests pass. All 5 audit papers verified. §9.11
2026-06-28 Gate 5 blind audit + pre-ref tail zone fix Gate 5: 24YKLTHQ (13p) + 4KCHGV2Z (9p) rebuilt post-P1. Found pre-ref body pages misclassified as tail_nonref_hold_zone. Root cause: _apply_zone_labels re-applies stale region_bus after ref partition. Fixed by stripping pre_ref block IDs from tail zone. 4KCHGV2Z P7: tail=20 → body=2+disp=5. All 286 figure/backmatter tests green. §9.12
2026-06-28 Gate 5 frontmatter fix series (3 fixes) 24YKLTHQ: author byline lowercase guard, metadata body_start fix, frontmatter heading normalization (no text matching). All 3 fixes verified on real paper, 461 tests pass, 0 new regressions. §9.13
2026-06-28 Test fix session: 25 tests reconciled Fixed 10 stale test issues: non_body_insert guard, caffard abstract, legend_like role, structural gate, backmatter heading render, state machine (done_degraded), body_zone anchor, rebuild backfill skip, truth surface docs, trace-vs-expectations (10 assertions). 616 OCR tests, 0 failed. Expectations updated for post-P1 behavior. §9.14
2026-06-28 Data-driven truth audit (2 papers) 2HEUD5P9 (27p) + 4AG67PBH (25p) — no vision (model limit). Found 3 pipeline defect patterns: zone_leak_frontmatter_to_body (2 papers), reference_boundary_body_mix (2 papers), title_repeat_page2 (1 paper). 12 ghost unknown_structural blocks in 2HEUD5P9. Findings saved to audit/2026-06-28-data-audit-findings.json. §9.15
2026-06-28 P0 author bio detection implementation Created ocr_bio.py with category-weighted bio scoring, Pass C (post_ref_bio_cleanup), figure match guards. Wired author_bio_asset role contract + pipeline. 30 new tests pass. 1041 total OCR tests, 0 regressions. Commits: e2f0c8a, 7810eb1. §9.16
2026-06-28 P1 author bio detection implementation Added residual_author_bio_pass (figure-residual portrait assets), extended post_ref_bio_cleanup for figure_caption, tag_figure_contained_text protection. 7 new P1 tests. 1018 total OCR tests, 0 regressions. Commit: 7a1cc5e. §9.17
2026-07-01 Audit fix commits + orientation-aware rotated figure normalization Commit 1 (2d40ad9) + Commit 2 (21bdfd0) + Commit 4 (7670227) landed, then refactored rotated-caption handling out of synthetic fallback into normal figure pre-match. Added PyMuPDF dir/wmode capture, same-page rotated settlement, rotated crop render. U746UJ7G now matches via same_page_rotated; KUR9PBJC unchanged. 422 regression tests pass. §9.18
2026-07-01 Asset-internal figure number recovery implementation Added extract_pdf_lines_normalized helper, _recover_missing_figure_numbers_from_assets pass in build_figure_inventory, 5 gate functions, 2 pattern constants. U746UJ7G figure_unknown_000 → figure_002 with recovered label "Plot of Criteria Time". 6 new tests. 428 regression tests pass. §9.19
2026-07-02 Round 2 truth audit + 3 targeted bug fixes for 37LK5T97 Batch-audited 10 fresh papers (5 GREEN / 4 YELLOW / 1 RED). 37LK5T97 found with Figure 1 broken (sidecar caption demoted) + 6 unmatched rotated tables. Fixes: (1) _is_sidecar_candidate guard in candidate_resolution, (2) adjacent_x+y_overlap in score_table_match for rotated captions, (3) rotated table render bbox+270° correction. Also: rotated figure crop quality fix (4x zoom + coordinate normalization in _crop_asset_from_pdf). Commits: 59cd01a, bd3f3b6, 86e0d14. 428 regression tests pass. §9.20
2026-07-02 Zone/role robustness completion — Figure caption prefix recovery + inline table fix + table matching audit Figure caption prefix recovery from PDF text layer (_recover_figure_heading_prefix): 5S7UI34M 4→9 matched figures, 33→1 unmatched. Inline <table> HTML role fix: 650 blocks now table_html after rebuild (priority bug: raw_label=table fired before <table> check). Table matching audit: 620 remaining media_asset pending full rebuild. 585 figure/table/role tests pass. §9.21

9. Historical Detail Archive

Fixed records and verbose session logs preserved below. The archive is read-only reference — active work is tracked in sections 2-4 above.

9.1 Gold Fixture Expansion + Bug Fixes (2026-06-16)

Full-day debugging session across 8 gold papers. 98 bug annotations, 8 pipeline fixes (F1-F10).

Root cause categories identified:

  1. Frontmatter noise unrecognized (ISSN, journal citation) → body pollution
  2. Cross-page text fragmentation
  3. Tail zone: body prose incorrectly converted to backmatter
  4. Backmatter heading gaps
  5. Heading merge/split logic
  6. Heading rescue from unknown_structural
  7. Heading prefix by role (not font size)
  8. Permissive figure matching
  9. All same-page assets included in group match
  10. Composite region text requirement removed

Fixes applied: ocr_roles.py, ocr_document.py, ocr_blocks.py, ocr_render.py, ocr_scores.py, ocr_figures.py, ocr.py

Layout-first Phase 1 pass added — table inventory relaxed for media_asset blocks, _should_keep_formal_caption_seed() added, tests/test_ocr_layout_first_regressions.py created.

9.2 Boundary Close-Out Pass (2026-06-17)

Plan: docs/superpowers/plans/2026-06-17-ocr-v2-closeout-single-plan.md

Tasks 1-5 executed:

  • Same-page reference/body boundary split (block-level vertical split by heading position)
  • False tail backmatter conversion reduction
  • Page-1 correspondence line → frontmatter_support
  • Preproof page-1 frontmatter: first-surviving-page logic, margin-band watermark detection, figure inner label extension
  • Active truth-file cleanup: reconciled P0-P2 close-out

Result: 202P/1F/43S (sole failure = pre-existing DW figure ownership)

Key commits: 6f68bf2, b7d369e, c7a9c93, 827a2cc, 9329843

9.3 Readiness-Gates (2026-06-18 ~ 06-19)

Plan: docs/superpowers/plans/2026-06-18-ocr-v2-readiness-master-plan.md

Gates implemented:

  • Gate 1 (Completeness): _summarize_page_text_coverage() + _classify_region_text_completeness() + audit_rendered_text_coverage()
  • Gate 2 (Figure ownership): 8 tasks — previous-page sequential fallback, DW Fig 3 xfail→pass, sidecar partition, fallback tightening. Final: 249/249 pass
  • Gate 3 (Ordering): _ref_number_sort_key regex extended, reference boundary normalizers
  • Gate 4 (Layout coverage): audit/coverage_ledger.json with readiness-class taxonomy, contract tests enforce named representatives

Blind audit (5 unseen papers): ALL PASS. Papers: 8VB9ZVQG, U746UJ7G, L6ALWJFP, PZ8B59K4, GU9R8EPE. No new failure families discovered.

Remaining residuals: ~40 stale audit truth blocks, ~50 edge-case misclassifications (low severity).

9.4 Cross-Page Caption Consumption Fix (2026-06-22)

Paper: SAN9AYVR Figure 24 double-emit. Cross-page figure ownership repair — reader/render chain consumed caption blocks on asset page instead legend page.

Fixes applied:

  • Reader payload records cross-page caption consumption on legend page
  • Render path suppresses original caption block on asset page
  • _recompute_final_unmatched_assets() added — orphan truth now matches final ownership truth
  • Cross-page duplicate ![[render/figures/figure_N.md]] eliminated

Test status: 150 passed (figures + reader + render)

9.5 Figure Merge Refactor + Visual Grammar Hardening (2026-06-21 ~ 06-23)

Core change: Replaced greedy region-growth with global distance clustering (union-find):

  • _cluster_page_assets(): horizontal <12% pw, vertical <8% ph, text-separator-aware
  • Caption-as-boundary: each legend claims assets by y-band
  • Composite parent detection: _build_composite_parent_figure_groups_visual_only()
  • Dense page arbitration: _build_dense_composite_parent_candidates() when ≥4 visual fragments
  • Ownership registry: FigureOwnershipRegistry with conflict detection
  • Figure/table separation: asset_family_hint + table-like veto (confidence ≥0.70)
  • Dedup refinement: _normalized_caption_body — internal punctuation preserved, terminal punctuation stripped
  • _ref_number_sort_key with [N] bracket support

Key fix (merge-gate closeout): Same-number-distinct dedup + grid collapse fix (highest-score selection, not first-match)

Test status: 216→225→261 passed (figure stack)

9.6 Rebuild Production Run (2026-06-26)

Full ocr rebuild --all (699 papers):

  • --resume checkpoint support (interrupted rebuilds resume)
  • ASCII tqdm progress bar fix (Windows)
  • Removed rogue [DEBUG] print in ocr_render.py:1295

Fixes found:

  • N6XCZD25: Body paragraphs misclassified structured_insert → body_spine_match flag in body_zone
  • Chinese Windows encoding: glob fallback in resolve_pdf_path; garbled meta.json auto-corrected on rebuild

9.7 Figure Number Inference + Container Admission (2026-06-26)

Figure number inference (leading [1] gap): 8-step algorithm in _infer_missing_main_figure_numbers(). N6XCZD25 figure_unknown_005 → figure_001.

Container admission rewrite (evidence-driven):

  • 7-method container extraction: page-sized/crop-like excluded, line-like→grouping only, vertical component merge
  • Three-phase per-page loop replaces lazy-cache
  • Blue sidebar box score 0.45→0.80 → rendered as [!NOTE] callout

Key commits: 45cf65e, bab0167, 5ba1a6d

9.8 UI Polish (2026-06-26)

Plugin dashboard cleanup:

  • Vercel-style CSS (cards, collapsible headers, status grid, issue summary funnel)
  • Title tooltip with title_full field
  • Table sorted by time descending
  • Click-to-copy on paper paths (pf-copy)
  • Global "Start Working" cleanup — Doctor/Repair only in Issues section
  • Run OCR button shows pending count
  • Redo OCR → Maintenance button (opens settings→maintenance tab)
  • DESIGN.md reference file added

9.9 Rebuild Audit + Index Repair (2026-06-27)

6 bugs fixed:

Issue Severity Status
sync_service.py: _time UnboundLocalError HARD BLOCK FIXED
Workspace fulltext never synced from OCR output HIGH FIXED
build_index early-return skips _build_entry HIGH FIXED
Field registry missing 9 common fields → 6000+ WARN MEDIUM FIXED
"700 papers missing fulltext" — false alarm MISLEADING NOTED
Base legacy fields — user choice INFO DEFERRED

Field registry additions: aliases, tags, journal, first_author, pmid, impact_factor, abstract, keywords, ocr_time.

9.10 Deep Investigation — 5 Fix Spec (2026-06-27)

5-paper vision audit: 25K5KZAQ, NC66N4Q3, 9TW98JH8, YGH7VEX6, XD2BPCMG.

Fixes identified and resolved:

Fix Issue Root Cause Resolution
Fix 1 Figure-internal text containment No spatial containment in build_figure_inventory Render-hygiene pass: 6 helpers + 19 tests ✅
Fix 2 Reference sorting [N] bracket gap regex only matches N./N), not [N] Two capture groups `r"^\s*(?:\d+
Fix 3 Backmatter boundary (CRediT/Ethics) 7.97pt < 11pt threshold; no container keywords Ref-anchored partition design 🔄
Fix 4 Figure caption → non_body_insert figure_caption in _INSERT_CANDIDATE_ROLES Removed list ✅
Fix 5 Demoted body paragraph in figure legends body_paragraph re-enters matching via legend detection Filter before _is_validation_first_legend_candidate ✅

Spec: docs/superpowers/specs/2026-06-27-figure-containment-and-backmatter-boundary-design.md Plan: docs/superpowers/plans/2026-06-27-figure-containment-implementation-plan.md

9.18 10-Paper Truth Audit + GPT Cross-Validation (2026-07-01)

Scope: 10 papers sampled from vault (3 green, 4 yellow, 3 red), batch-audited via ocr_truth_audit.py + annotated page vision agents.
1 complete vision audit: U746UJ7G — 0 matched figures, root cause = vector-rendered Figure 1+2 (52 drawing paths, 0 embedded images).
Key bugs found:

  • _TABLE_PREFIX_PATTERN 3 处只认 \d+，不认罗马数字 I/II/III（KUR9PBJC pages 4/5/7）
  • raw_label=figure_title 的 Table N caption 无 guard → 进 figure_caption
  • vision_footnote 里 "This figure..." 被吞成 footnote（U746UJ7G p8:4）
  • unmatched_assets 重复计数（matched 后又在 unmatched 列表）
  • HTML table block 断连（2YW2MJBL）
  • Vector figure pipeline 盲区（U746UJ7G）

False positives: reference_span_error ~90% noise, same_page_boundary_error 100% unreliable.

GPT 二审修正 8 处: Roman+S prefix 三处同步、score hard gate、tie-breaking、dedup、_is_near_figure_media helper。
Outputs:

  • Spec: docs/superpowers/specs/2026-07-01-ocr-audit-findings-for-gpt.md
  • Plan: docs/superpowers/plans/2026-07-01-ocr-audit-gpt-fix-plan.md
  • Per-paper vision reports: local://*-vision-report.md Execution: 3/4 commits implemented (2d40ad9, 21bdfd0, 7670227). Commit 3 (unmatched dedup) pre-existing from PR3 (4ab227e). Follow-up refactor moved rotated-caption handling earlier: PyMuPDF dir/wmode now preserved in span_metadata; rotated vision_footnote candidates enter normal legend matching; same_page_rotated matches carry rotation_correction_deg into crop/render. U746UJ7G now produces a normal matched figure (figure_unknown_000, settlement_type=same_page_rotated) instead of synthetic fallback. 422 regression tests pass.
  • Block reviews: audit/<KEY>/block_review.jsonl（U746UJ7G, KUR9PBJC, CGGYTEEQ, 7FNV9AW2 etc.）

9.19 Asset-Internal Figure Number Recovery (2026-07-01)

Problem: U746UJ7G rotated figure — "Figure 2. Plot of Criteria Time..." was absorbed into chart asset bbox during OCR assembly. The rotated prematch refactor (9.18) produced a matched figure (figure_unknown_000) but without figure_number. No way to assign Figure 2 to the figure.

Solution: Metadata-only recovery pass after synthetic vector fallback, before dedup.

  • extract_pdf_lines_normalized in ocr_pdf_spans.py extracts PDF rawdict text lines page-by-page, normalizing coordinates to OCR space
  • _recover_missing_figure_numbers_from_assets iterates matched_figures without numbers, scans each matched asset's bbox for internal PDF line labels ("Figure N.", "Fig. N:")
  • _needs_asset_internal_figure_number_recovery: gates by figure_id prefix (figure_unknown_* or synthetic_figure_*) + text description signal
  • _looks_like_internal_figure_label: rejects full-sentence patterns ("Figure X shows...") via regex match
  • _asset_edge_band_score: geometric rejection for lines in asset center (not label position) or covering >15% of asset area
  • Coordinate normalization is caller responsibility (done in ocr_rebuild.py)

Files changed:

  • paperforge/worker/ocr_figures.py: 7 new functions + 2 pattern constants + signature change to build_figure_inventory + recovery pass call
  • paperforge/worker/ocr_pdf_spans.py: extract_pdf_lines_normalized helper
  • paperforge/worker/ocr_rebuild.py: call helper, pass page_pdf_lines_by_page to inventory builder
  • tests/test_ocr_figures.py: 6 new tests (basic recovery, duplicate rejection, normal-fig unaffected, multi-label conflict, center rejection, overlap gate)

Spec: docs/superpowers/plans/2026-07-01-ocr-asset-internal-figure-number-recovery-plan.md Execution:

  • U746UJ7G verified: figure_number == 2, figure_id == "figure_002", recovered_label_text contains "Plot of Criteria Time", flags contain figure_number_recovered_from_asset_text
  • 428 regression tests pass (422 existing + 6 new)

9.20 Round 2 Truth Audit + 37LK5T97 Bug Fixes (2026-07-02)

Scope: 10 new papers batch-audited via ocr_truth_audit.py (high-risk mode). 5 GREEN / 4 YELLOW / 1 RED.
RED paper: 37LK5T97 — "Both IM and EC ossification occurs during the bone-healing process"

Three bugs found and fixed:

  1. Figure 1 sidecar demotion: Caption (left column, 246px) and image (right column, 693px) had zero x_overlap. _is_near_figure_media() missed it, _looks_like_figure_narrative_prose() caught the long description, and candidate_resolution demoted to body_paragraph. Fix: Added _is_sidecar_candidate() guard — checks vertical overlap with media_asset when horizontal overlap is absent. Caption preserved as figure_caption_candidate.

    • File: paperforge/worker/ocr_document.py
    • Result: Figure 1 matched as figure_001 (caption block_id=2, asset block_id=9)

  2. Rotated table caption matching: Tables 1-3 had rotated captions (span dir=[0,-1], vertical text beside table body). score_table_match required x_overlap + asset_below_caption, both failed for rotated sidecar layout. Fix: Added adjacent_x + y_overlap_with_asset scoring branch when caption has rotated text and x_overlap < 0.5.

    • File: paperforge/worker/ocr_scores.py
    • Result: All 6 tables matched (has_asset=true, score >= 0.65)

  3. Rotated table render orientation: Table body also had dir=[0,-1] (rotated 90° content on portrait page). Rendered JPEG showed vertical text. Fix: Added _table_has_rotated_content() helper; computes union render_bbox (caption+asset) and render_rotation_deg=270 in table entry; render loop passes rotation_deg to _crop_asset_from_pdf.

    • Files: paperforge/worker/ocr_tables.py, paperforge/worker/ocr_objects.py
    • Result: Tables 1-5 rendered at correct orientation (1908×2858 → 2858×1908)

Additional quality fix: Rotated figure crop quality improved in prior session — Matrix(2,2) → Matrix(4,4), PIL rotation from pix.tobytes("png") (single pass, no double JPEG), and OCR→PDF coordinate conversion fix. U746UJ7G figure_002: 2618×1914 px, 5.0M px (4.5x improvement).

Commits: 59cd01a (figure quality+rotation coord fix), bd3f3b6 (sidecar+table match), 86e0d14 (table render rotation)
Tests: 428 regression tests pass (no new tests added, existing cover the affected paths)
Analysis: docs/superpowers/analysis/2026-07-02-37lk5t97-figure1-sidecar-bug-analysis.md
Audit findings: docs/superpowers/specs/2026-07-02-ocr-truth-audit-round2-findings.md

9.21 Figure Caption Prefix Recovery + Inline Table Fix (2026-07-02)

Problem 1: PaddleOCR fails to detect standalone "Figure N" / "FIGURE N" headings in bold/small-caps fonts. Caption body is captured as figure_caption_candidate but lacks the figure number prefix → _is_formal_legend() fails → caption never enters matching pool.

Fix: _recover_figure_heading_prefix() in ocr_figures.py checks the PDF text layer (via existing page_pdf_lines_by_page infrastructure — no extra PDF open) for "Figure N" lines. If the next PDF line (by y-order) shares ≥15 common-prefix chars with the OCR caption text, the heading is prepended. Runs BEFORE the zone/style filter so recovered captions pass through to legend matching.

Result: 5S7UI34M (PVA综述): 4→9 matched figures, 33→1 unmatched (p1 logo). HQAQBSBP: Figure 5 recovered. 372 figure tests pass.

Problem 2: Inline <table> HTML blocks with raw_label=table hit the raw_label="table" → media_asset fallback (ocr_roles.py:1355) before reaching the <table> → table_html check (line 1456, gated by raw_label="text"). Plus ocr_document.py:6120-6121 converted table_html to table_html_candidate — a role with no downstream handler → structural gate downgraded to unknown_structural.

Fix: (1) Moved inline <table> check before raw_label=table fallback in assign_block_role(). (2) Added table_html verifier in structural gate (self-identifying, accepts if text starts with <table>). (3) Removed dead table_html → table_html_candidate conversion.

Result: AH6Q7DLC (worst case, 30 blocks): 29/30 → table_html (1 = reference_item). 585 figure/table/role tests pass.

Files changed:

  • paperforge/worker/ocr_figures.py — _recover_figure_heading_prefix() + body_zone filter guard
  • paperforge/worker/ocr_roles.py — inline <table> before raw_label=table
  • paperforge/worker/ocr_document.py — removed table_html_candidate dead path
  • paperforge/worker/ocr_structural_gate.py — table_html verifier