feat(sentry): wire the diagnostic metric emitters for shutdown, IPC errors, and telemetry failures

[gmaclennan]

Closes #190.

backend/lib/metrics.js shipped three emitters with no production call sites, so their metrics never reached Sentry. This wires them.

shutdownPhase now records via sentry.withSpan, which routes the always-on phase metric by op prefix (shutdown.* → comapeo.shutdown.phase_duration_ms, everything else → the boot metric as before). The backend's control-socket shutdown handler wraps its close steps in three spans — shutdown.close-servers (control + fastify + RPC in parallel), shutdown.close-manager, and shutdown.close-map-server — mirroring how the boot phases record and matching the three-phase cardinality budget in the plan doc.

ipcError records comapeo.ipc.errors where socket errors surface: the messageerror listeners on the control socket (SimpleRpcServer) and the comapeo socket (ComapeoRpc), plus the broadcast catch when a client postMessage throws.

telemetryForwardingFailure records comapeo.telemetry.forwarding_failures at the envelope sink's three failure points in backend/lib/sentry.js: the forwarding transport when the sink throws, the pre-listen queue when it overflows and drops a frame, and the setSink drain when the sink throws.

Each emitter gets a test through its real call path: withSpan("shutdown.*") and a throwing sink in sentry.test.mjs (against the real SDK), and a malformed frame on each socket in simple-rpc.test.mjs / comapeo-rpc.test.mjs. The two #190 "not yet wired" markers are removed from the §9.2 tier table; the sync rows' markers stay — #80 owns those.

Verified with npm run lint, npm run build, npm run test (69 passed), and npm run backend:test (77 passed).