The indent trigger was using .abs() which fired on both increased indent
(non-indented → indented) AND decreased indent (indented → non-indented).
This caused drop-cap style paragraphs (indented first line, flush-left
continuation) to incorrectly split into two blocks.
Per plan Phase 4.4 heuristic #2, indent change should only trigger when the
current line is MORE indented (to the right, larger x0) than the block
average - i.e., a new paragraph starting after non-indented text. It should
NOT trigger for decreased indent (first line indented, rest flush-left).
Fix: Remove .abs() and only check if line_x0 - block_avg_x0 > threshold.
Tests:
- test_indented_first_line_new_block: PASS (non-indented → indented splits)
- test_indented_first_line_of_paragraph_not_split: PASS (drop cap stays together)
- All 179 line module tests: PASS
The bead description mentioned compile errors in hash.rs from API drift,
but those errors were either already fixed or misattributed. The API usage
was already correct:
- compute_fingerprint already takes 3 arguments with source
- len() already propagates Result with ?
- read_at method already used correctly
- Catalog fields accessed via trailer correctly
Only cleanup: removed unused std::fs::File and std::io imports.
Verification: notes/bf-4mkhv.md
The native PyO3 module returns raw dicts via pythonize, but the Python SDK
API expects typed dataclass objects (Document, Page, Metadata, etc.) to be
consistent with the subprocess fallback and test expectations.
Updated wrapper functions in __init__.py to convert native results:
- extract(): wraps dict in Document.from_dict()
- extract_stream(): wraps yielded page dicts in Page.from_dict()
- get_metadata(): wraps dict in Metadata()
- hash(): wraps string in Fingerprint.from_string()
- classify(): wraps dict in Classification()
- search(): wraps yielded match dicts in Match
The native PyO3 entry points (extract, extract_text, extract_stream) were
already implemented with:
- extract: uses extract_pdf + pythonize for PyDict conversion
- extract_text: uses extract_text for plain String return
- extract_stream: uses extract_pdf_streaming with custom StreamIterator
All kwargs parsing with strict validation (unknown kwargs raise TypeError)
was already in place.
Acceptance criteria:
- pdftract.extract() returns Document object with pages/metadata
- pdftract.extract_text() returns plain text string
- pdftract.extract_stream() yields Page objects
- Unknown kwarg raises TypeError
The extract_markdown stub was calling extract_text instead of
extract_text_fn, causing a compilation error. This fixes the
function name to match the exported function from extract_text.rs.
This completes the extract_text PyO3 entry point implementation,
which was already present in extract_text.rs and lib.rs.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add CI validation script for checking unauthorized expose_secret() call
sites. The script validates that all uses of expose_secret() are in
approved locations (SecretFingerprint and test code).
Also add verification note summarizing the bead completion status.
Per pdftract-5l9m acceptance criteria:
- CI grep guard rejects unauthorized expose_secret() call sites
- Verification documents existing SecretString wrapping status
Co-Authored-By: Claude Code <noreply@anthropic.com>
Fixed scripts/check-provenance.sh to properly validate PROVENANCE.md
against actual fixture files. The script was failing silently due to
subshell EXIT trap removing temp files before parent could read them,
and arithmetic expansion returning exit code 1 on zero value.
Changes:
- Replaced subshell pipes with process substitution
- Moved temp file cleanup to after reading
- Added validated variable initialization
- Added || true to prevent exit on zero arithmetic
All 200 classifier corpus fixtures have valid provenance entries
with matching SHA256 hashes. PROVENANCE.md already existed with
complete documentation.
Refs: pdftract-5z5d8
Co-Authored-By: Claude Code <noreply@anthropic.com>
- Create tests/fixtures/classifier/ with 200 synthetic PDFs:
- 50 invoices with bill-to/ship-to, item tables, totals
- 50 scientific papers with abstracts, sections, references
- 50 contracts with clauses, legal terminology, signatures
- 50 misc documents (8 receipts, 8 forms, 7 bank statements,
7 slide decks, 7 legal filings, 6 book excerpts, 7 magazines)
- Add MANIFEST.tsv mapping each document to its expected type
with source URL and license (all MIT-0 synthetic data)
- Add scripts/generate_test_corpus.py to regenerate the corpus
using reportlab for PDF generation
- Add tests/test_classifier_corpus.rs with validation harness:
- test_corpus_manifest_validity: verifies manifest structure
and file existence (PASSES)
- test_classifier_corpus_accuracy: will validate precision/
recall/F1 when classifier is implemented (SKIP for now)
- test_classifier_reproducibility: will verify deterministic
classification (SKIP for now)
- Add tests/fixtures/classifier/README.md documenting corpus
structure, generation process, and acceptance criteria
Total corpus size: ~0.4 MB (each PDF < 5 KB)
Acceptance criteria (from plan.md Phase 5.6):
- Per-class precision and recall >= 0.85
- Macro-F1 >= 0.88
- Reproducibility: identical output for same document
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
