Pathfinder retrieval upgrade + query-source observability

[jpr5]

Summary

• Repoint docs indexing to the live shell-docs tree. Pathfinder was indexing the retired docs/content/docs/ tree; the docs source now points at showcase/shell-docs/src/content/docs/ (path, URL derivation, webhook path-triggers) so search returns current pages.
• Inline MDX @/snippets/* imports before chunking so snippet-composed pages index with real content instead of near-empty stubs.
• Embed chunk title + heading path into the vector text so precise symbol/prop/heading queries keep their strongest anchor (plus an embedding count-mismatch guard and a metadata-precedence fix).
• Hybrid (vector + keyword) search with min_score: 0.3 on all four tools; scope code/ag-ui-code indexes (exclude examples/showcase/.next/*.d.ts/generated/proto; broaden ag-ui-code to the Python + community SDKs); sharpen tool descriptions.
• Query-log observability: tag each row with request_source (X-Pathfinder-Source) + session_id, and add a low-confidence windowed metric so our own analysis traffic is separable from real users.

Why

From a 30-day analytics gap analysis: the top empty-result drivers were the indexer pointing at the retired docs tree, snippet-composed pages indexing empty, and no way to distinguish synthetic/analysis traffic from real users.

Test plan

• tsc --noEmit clean; prettier --check "src/**/*.ts" clean
• full suite green (3608 tests / 242 files)
• CI green on this PR

Follow-up (separate PR)

Pre-existing indexing-durability + dashboard-consistency issues surfaced during review (delete-before-upsert chunk loss, "All time" 366-day cap, ag-ui-docs .md 404s, code-fence heading-path attribution) will be handled in a dedicated PR — none are introduced here.

🤖 Generated with Claude Code