Sentry: trace requests only, MCP calls at a lower rate; KG writes each edge once per pass

[keysersoft]

At 5 % the cloud backend sent about 700,000 spans a day to Sentry, far past a plan's quota. By volume:

• database spans of the knowledge-graph ingest, attached to the MCP call that scheduled it: a SELECT and an UPDATE on kg_edges for every shared value, even when many values link the same two nodes;
• root spans with no request behind them (pg-pool.connect, prisma:client:*, bare SQL), each sampled as a transaction of its own.

Changes

• common/sentry-sampler.ts: only HTTP requests are traced. MCP routes use SENTRY_MCP_TRACES_SAMPLE_RATE, default a fifth of SENTRY_TRACES_SAMPLE_RATE. /health is never traced. Incoming sampling decisions are not inherited, because MCP clients send their own sampled traceparent.
• KgObservationalService.ingestOrganization runs under Sentry.suppressTracing. Errors are still reported.
• correlate() counts bumps per distinct edge and writes each edge once. The end state is the same as the sequential writes: observations, confidence, the last matchKey set, and status (a rejected or manual edge keeps its own). The references confirmation runs once per node pair.

Tests

• New sentry-sampler.spec.ts.
• New KG test: four values linking the same nodes give one read and one write per edge, observations +4, and the same confidence as four sequential bumps.
• Backend suite passes locally, except the HANA suite, which needs the hdb driver that isn't installed here.
• Typecheck and lint are clean.