Knowledge graph: typed edges, junk-link suppression, and graph health stats

[rahilp]

What

Turns the knowledge graph into a real typed graph instead of 94% generic relates_to edges, per the amended v3 plan.

• Typed edge inference at capture time: follows, caused_by, decided drawn in addition to generic relates_to (junk-link suppression prevents near-duplicate confusion from creating misleading connections).
• Update and merge paths re-infer edges so the graph stays current when content changes; typed edges replace only inferred relates_to pairs, never explicit links.
• Nightly backfill can emit follows when an entry's kind is already classified; refuses long-mirrored sources.
• Insight calls parse a relationship + direction before the insight gate, retiring only inferred relates_to edges.
• GET /stats/graph (admin only, gated behind ?deep=1, unscheduled) reports graph health with gap distribution and EXPLAIN-planned query assertions.
• Weekly sweep instead of nightly (~7x cheaper amortized).
• MCP link tool description now explains each edge type and direction; kind gating wired into POST /link and MCP link.

Constraints honored

• Brain-agnostic, zero added AI/LLM calls, no recall-scoring changes, free-tier D1/Workers AI/subrequest budgets respected (edge writes batched 5→2 subrequests).

Verification

• 3379 tests passing (from 3266 baseline, +113)
• tsc --noEmit clean; check:scope 99 queries / 53 exceptions
• ADR/decisions: ?deep=1 stays manual/unscheduled (no edges(workspace_id) index, user decision; ~5M row-visit cost documented)