chat: log the undelivered-turn cause and thread a reportError refId (CL-6644, part A)

[TheGreatAxios]

Summary

CL-6644 part A: diagnosability. A turn that fails end-to-end left zero log
trace — dispatchTurnBatch's catch logged through a bare getLogger call,
never through @corbits/error-sink's reportError, so an undelivered
notice never carried a refId a person could quote to support.

• dispatchTurnBatch's catch now calls reportError (operation
  chat.dispatchTurn, tenant/room/agent context, the message ids) before
  logging and posting the notice, and threads the returned refId into
  postUndeliveredNotice.
• The notice text now ends with (ref <refId>) — cheap enough to fold in
  now rather than waiting on CL-6641 item 3's own UI surface.
• Added @corbits/error-sink as an explicit packages/chat dependency.

Part B (wake never completes) — not fixed here

I spent the session's timebox tracing the wake hop
(ensureAwake/wakeByAddress/wakeFoldedRun/deployAdoptedWorkflowFromSource)
and could not pin the stalling hop to a specific line with the confidence
this repo's "no guessing" bar requires, without a live repro run against
the shared stack (which I was told not to restart unless dead). The
strongest lead: c83f3821 ("Sidecar boot restore: skip parked
deployments...") justifies skipping a parked record's boot restore by
saying it "resumes correctly on the next message... the same wake path
that already handles a cold deployment" — but 0fd3fbc8 ("replace idle
park/wake with reap-and-relaunch teardown") had already deleted that
in-place wake-on-mail-route path days earlier. wakeByAddress
(packages/chat/src/platform-adapter.ts) still carries a CL-6267 comment
describing that same deleted mechanism ("the sidecar's own park
wake-handler owns respawning a parked-but-still-announced deployment").
This mismatch is worth a live-repro session to confirm or rule out before
touching sidecar code — filing as a Linear comment on CL-6644 rather than
guessing a fix blind.

Test plan

• WORKBENCH_CHECK_SINCE=origin/main bun run typecheck — pass
• WORKBENCH_CHECK_SINCE=origin/main bun run lint — pass
• WORKBENCH_CHECK_SINCE=origin/main bun run test — pass, including a
  new red/green test asserting the notice carries (ref ...)
• Live repro (send → typing → real Ollama reply) — not run; see Part B
  note above