Remove the closed PDF backend study

[willhea]

What

Removes the closed PDF backend study from the tree, under the research-retention rule (AGENTS.md, "Research artifacts are working material"). #740 closed it as inconclusive, with no architecture recommendation. Git history keeps everything; CLOSEOUT.md at 4171e32d is the citable record.

Removed	Size
docs/research/pdf-backend-bakeoff/	bulk of the ~27 MB
docs/research/staffer-delivery/	15 files
tests/data/p3/ (3 PDFs)	read only by bake-off probes (confirm_safe_failure.py, x01_contamination.py)
.gitignore bake-off block	holdout corpus and external-validity result rules

328 tracked files deleted. CLOSEOUT.md goes too: once its conclusions have homes, it is a closure document whose only job is recording what was deleted.

What survives, and why

Two probes still answer live questions, so they move to scripts/ instead of being deleted:

Script	Live question	Verified on develop d3935ee1
scripts/pyodide_parity.py + .mjs (was staffer-delivery/probes/verify_parity.py, parity_pyodide.mjs)	The gate #751 names for the lazy pypdfium2 import; in-browser delivery is still open in #112	3 fixtures × 2 artifacts IDENTICAL, exit 0. --mutate: every HTML row MISMATCH, exit 0. No node dir: exit 2. Hashes identical before and after the move
scripts/make_scan_fixture.py (was pdf-backend-bakeoff/probes/make_scan_fixture.py)	The only executable reproduction of open bug #550 (two scanned PDFs compare as "no changes"); no test owns it	DEFECT REPRODUCED: both 20-page scans extract 20 lines, guard_declines=False, changes=0

Both are cataloged in scripts/README.md. make_scan_fixture.py needs pypdf, which only the probes' own requirements.txt provided. It runs with uv run --with pypdf rather than adding a project dependency, and its catalog row says to delete it once a test owns #550.

Fixed in passing: verify_parity.py's docstring said --mutate exits 1; the code exits 0 when the corruption is detected and 1 when it is not. The harness no longer defaults to node_modules beside itself, since that would land untracked in scripts/.

Where the conclusions went

Written as present-state facts into the ADRs that own them (ADRs here are living records), each with one evidence link to CLOSEOUT.md at 4171e32d:

• ADR 0002 (PDFium engine): PDFium is the engine for anchoring and chrome handling, not for a measured accuracy lead. No backend dominates (pdfminer.six best against the XML, PDFium-WASM exact on production output and faster), and no design for plugging in a second backend was selected.
• ADR 0003 (client-side extraction): the Python engine runs under Pyodide with byte-identical XML output; the module-scope pypdfium2 import is its one obstacle (The XML comparison cannot load in a browser because it imports a PDF library it never uses #751); PDFium-WASM exposes the per-glyph data the extractor uses. The PDF path has not run end to end in a browser.
• ADR 0011 (local-only processing): CSP alone cannot give a browser channel zero egress. window.open and WebRTC are outside CSP; Speculation Rules is closed only when script-src drops 'unsafe-inline'. Closing the rest needs a browser- or device-level control.

pyproject.toml's extend-exclude stays, since provision-matching/ and pdf-matching-convergence/ still have probes and other studies have write-ups. Its comment loses the clause about a frozen preservation manifest, which existed only in the bake-off.

Consumer trace

• git grep for both directory names, data/p3, the fixture names, and the probe and module names: only .gitignore outside the study, plus the new evidence permalinks.
• CI workflows, pytest config (testpaths = ["tests"]), tests/test_research_probes.py (covers provision-matching only), Dependabot config: no references. The pip bumps to pdf-backend-bakeoff/probes (build(deps): bump pypdf from 6.16.1 to 6.19.0 in /docs/research/pdf-backend-bakeoff/probes #745, build(deps): bump pypdf from 6.15.0 to 6.16.1 in /docs/research/pdf-backend-bakeoff/probes #717, build(deps): bump pypdf from 6.14.2 to 6.15.0 in /docs/research/pdf-backend-bakeoff/probes #566) came from the probes' requirements.txt, which this removes.
• No test globs into tests/data/ recursively; sweep_bill_dirs covers the corpus trees only.
• Open issues and PRs searched; no open PR touches the removed paths. Issue references are listed below.

Issue text updated for the new paths

Applied with the maintainer's approval, in place (all three are the maintainer's own text):

• The XML comparison cannot load in a browser because it imports a PDF library it never uses #751 body: Verification now names scripts/pyodide_parity.py --node-dir <dir>; the study README link is a permalink at 4171e32d.
• Two different scanned PDFs compare as "no changes" instead of being declined #550 first comment: reproduction command is now uv run --with pypdf python scripts/make_scan_fixture.py --out /tmp/scans.
• Open decision: how DeltaTrack reaches staffers (delivery channel), waiting on Hill IT, workflow and packaging signals #112 comment: the RESULTS.md link is a permalink at 4171e32d.

The new scripts/ paths resolve once this merges.

Verification

• ruff check . and ruff format --check .: clean.
• Fast suite from outside the checkout (-m "not slow and not browser", as CI runs it): 2130 passed, 4 skipped, 15 xfailed.
• tests/test_adr_index.py and tests/test_docs_consistency.py pass with the ADR edits.
• The slow tiers were not run locally: nothing they read was removed or changed.

🤖 Generated with Claude Code