feat: admit enterprise-document evidence into query and packet

[1wgrumph]

Summary

Admits enterprise-document evidence into bran query and bran packet (#46). This connects the four adapters (DOCX, XLSX, PPTX, PDF) to the normal query surface.

• Discovery: documents are found by extension, capped at 64 files. Each adapter projection is wrapped in the RFC: BRAN evidence envelope for managed enterprise-document parsers #20 evidence envelope and schema-validated before admission. An envelope that cannot be represented is refused as oversized.
• Output: query and packet return admitted anchors in an additive document_evidence member, capped at 32 matches, with source type, native locator, source digest, fidelity and derivation. Document matches count toward query coverage and outcome. Packet excerpts obey the byte and token budgets and are included in the accounting.
• Safety:
  • OCR-derived anchors are labelled and never ranked as byte-derived text.
  • Refused or DLP-failing documents never enter a packet. They get a typed refusal that does not echo the unsafe path.
  • Locators themselves pass the DLP and public-boundary checks.
  • Document paths are excluded from Markdown ranking.
• Inspection: bran document inspect prints an envelope and writes nothing.
• Unchanged output: repositories without documents produce byte-identical output. bran-cli now depends on the in-workspace bran-document crate; bran-core still has no dependencies.

Closes #46

Verification

• ./tools/ci/check.sh --fast and --full pass on the branch, which is current with main. No new external dependencies.
• document_admission CLI tests: 15 pass, covering all four formats, inspection, refusals, DLP, budgets, coverage, determinism and unchanged document-free output.
• Independent review (Codex GPT-6 Astra) confirmed byte-identical document-free output on five repositories and the 64-file and 32-match caps. It returned REPAIR_REQUIRED with five findings, all fixed in f846fc1 with tests that failed first:
  • file names bypassed DLP;
  • refusals echoed unsafe paths;
  • document excerpts skipped the packet budget;
  • oversized envelopes were admitted despite failing the schema;
  • document matches reported the query as a miss.

Not covered: no real OCR producer exists here, so the OCR labelling is pinned by unit tests.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GXsmZAgLAftTb8E5oswz82