feat(classification): generated-code exclusion + PII policy enforcement (6.5, D5/G7)

[officialCodeWork]

Phase 6.5 — Generated/vendored classification & PII policy (failure modes D5, G7)

Closes the D5/G7 criteria of the Phase 6 gate.

Generated-code classification (D5)

Machine-generated code is retained in the graph as lineage / API metadata but excluded from match candidates — it is not authored UI, so a screenshot or ticket must never resolve to it.

• New isGeneratedFile classifies a file by:
  • a __generated__/ or generated/ directory segment, or a .generated. / .gen. filename infix;
  • an @generated / DO NOT EDIT / AUTO-GENERATED banner in the file head;
  • a sourcemap-less minified line (≥ 3000 chars).
• A post-pass tags component and hook nodes in those files with a generated flag (uniform across function components, class components, hooks).
• matchComponents drops generated components from the candidate pool and the IDF corpus; their nodes/edges — including data sources — stay for tracing.

PII policy (G7)

• New docs/security.md: screenshots are ephemeral (processed in memory, never written/embedded/logged); the graph and corrections store hold terms only, never images.
• Enforced by packages/vision/src/policy.test.ts — greps the vision source for filesystem-write / image-persistence APIs (writeFile*, appendFile*, createWriteStream, toDataURL, …) and asserts the package never imports fs. Fails CI on regression.

Fixture

d5-generated-code: path-classified GlyphCatalog + banner/filename-classified SchemaViewer + hand-written RevenuePanel. Both generated components trace to their endpoints yet decline on their rendered text; RevenuePanel matches and is not poisoned by generated text mixed into the query.

Verification

• pnpm eval → 314 pass · 0 fail · determinism 1.000 · all metrics 1.000
• 6 new parser-react tests + 7 vision policy tests; full suite, pnpm typecheck, pnpm lint green.

Tracker updated. 6.1 / 6.2 / 6.4 still to land for the full lifecycle phase.

🤖 Generated with Claude Code