fix(chat): follow-ups from the Copilot refresh-retention review (#2068) #2112

Follow-up to #2068 (LOW findings from the merge pass's reviews; none blocked the merge).

• Peer claims on the success path: a successful Copilot conversation read reuses the peer claims gathered at the start of the request (server/conversation.ts, success-path copilotTranscriptPath(..., resolved.peerClaims)). A peer that claims the same session during the read is seen only by the next request.
• Liveness is not identity in the failed-read recheck: the recheck fails closed when a cached peer Copilot process has exited (processAlive). It does not catch pid reuse within one herdr call, a zombie peer, a new Copilot in a peer pane that had none cached, or a process that changes identity without exiting. Comparing the process start time (as process-start.ts can) would close most of these.
• Unit coverage: no unit test covers a failed read with a live peer that still retains history (503); only the contract test does.
• Persistent unexpected read error: a read error that never clears keeps returning 503, so Chat keeps the old messages indefinitely instead of reaching the terminal fallback.
• Comment wording: server/conversation.ts says "never reuse peer claims here" just above code that reuses cached peer process info; "never reuse the request's earlier peer claims" says what is meant.
• Test coverage of the 409 late-response case: the 409 variant of the late-response deadline test passes on the base too, so it guards against a future regression but does not show this fix.