Summary

• Re-running codeiq index skips parse+detect for files whose content hash already lives in the cache (~5-10× faster on unchanged trees).
• Re-running codeiq enrich short-circuits when the cache manifest hash matches what the graph stored on its last run (86 ms vs 582 ms full rebuild on fixture-multi-lang).
• enrich is now safely re-runnable: a new Store.Reset() drops prior CodeNode+GraphMeta data before a full rebuild, so re-runs no longer collide on primary keys.
• New codeiq diff subcommand previews the cache vs disk delta as JSON without touching anything; index --force and enrich --force/--diff flags surface the new behaviour.
• Linker emissions (Topic / Entity / ModuleContainment) carry a source tag so future delta-apply work can wipe + re-emit them en bloc via Store.WipeLinkerEdges.

What's not in this PR (deferred)

• True per-file delta-apply to the graph (use ReplaceFile per changed path instead of full rebuild). The primitives ship here — Store.RemoveFile / InsertFile / ReplaceFile are wired and tested — but the enrich orchestrator still does a full rebuild when the manifest mismatches. The short-circuit case (the common one) is fully delta-aware.
• MCP-side freshness signals / auto-bootstrap — separate plans.

Phases (12 atomic commits)

Test plan

• CGO_ENABLED=1 go test ./... -count=1 — 937 passing across 44 packages
• CGO_ENABLED=1 go test ./internal/analyzer/ ./internal/graph/ ./internal/cache/ -race -count=1 — 143 passing, no data races
• go vet ./... — clean
• Manual smoke on testdata/fixture-minimal:
  • First codeiq index .: 5 files Added, 0 cache hits
  • Second codeiq index .: 5 Unchanged, 100% cache hits
  • Second codeiq enrich .: "enrich short-circuited: graph already matches cache manifest"
  • codeiq enrich --force .: rebuilds cleanly without PK collisions
  • codeiq diff .: returns JSON with empty Added/Modified/Deleted and 5 Unchanged
• New integration tests cover: incremental == full determinism, 3× idempotent re-runs short-circuiting, delete-then-recreate-with-same-content
• Reviewer: try on a real polyglot project to verify the short-circuit timing claim at scale

Known follow-up (out of scope)

govulncheck flags 2 stdlib net/http CVEs (go1.26.2 → fix in 1.26.3) — toolchain bump, separate work.