[lavaman131]

A build survived a hosted outage of 89 seconds with one second of its 90-second post-restart window left. A slow deploy or a host reboot would have ended running builds.

Server: a longer, configurable window

The server now waits 10 minutes by default for a build's machine to come back after a restart. That matches how long the connector keeps an agent running without reaching Chopin. CHOPIN_BUILD_RESUME_MS changes the window; it accepts 1000 to 3600000 ms and is documented in docs/self-hosting.md.

One rule decides whether a build may still resume, and every place that used the old 90-second window now asks it, so they never disagree:

Place	Follows the new window?	Why
Sweep (ending a build nobody carried on)	Yes	This is the window itself. A build whose machine never returns ends as Your local agent disconnected. once it passes.
unavailable (503) answers to the agent's run credential	Yes	The agent should keep retrying exactly as long as the build can still resume, and no longer.
Takeover of a forgotten connection's build by the machine's new connection	Yes	A machine back at minute 5 must be able to take the build back.
A machine that registered again without carrying its build on	Gets 1 minute, measured from its oldest live connection	A connector still running the build renews it within seconds of registering. It gets a heartbeat's allowance (the same 60 s a renewal grants) instead of the whole window, then the build ends as Your local agent stopped responding.
Connection expiry (90 s without hearing from a connector)	No	It's unrelated to restarts. Raising it would delay noticing a vanished machine in normal operation.
Connector's two-minute "forgotten" window	No	After a restart, a forgotten connection registers again on the next call; that window only covers repeated forgotten refusals.

Connector: retries are visible

While Chopin is out of reach during a run, the heartbeat loop prints a line at most every 10 seconds, then reports the outage length when contact returns:

  … Lost contact with Chopin. The agent keeps working; reports wait until it is back.
  … Still trying to reach Chopin (attempt 3, 12 s, ConnectionRefused), the run keeps going.
  … Still trying to reach Chopin (attempt 9, 36 s, HTTP 502), the run keeps going.
  ✓ Back in contact with Chopin after 44 s.

The line names only the HTTP status and/or the server's error code, or the network error code. It never includes the failure's message, so no token can appear. A test checks this with a bearer token placed in an error message.

The connector change needs a CLI release (no protocol change). There's a note under "Notes for the next release" in docs/releasing-cli.md.

Tests (written red first)

• routes.test.ts: a build carries on after a five-minute outage. The agent gets 503 unavailable every minute, then the machine's new connection takes the build back.
• routes.test.ts: a build whose machine never comes back ends as disconnected, both after the default 10-minute window and after a configured 2-minute window.
• routes.test.ts: a machine that comes back without its build ends it as stopped responding one minute later.
• config-build-resume.test.ts: the default, the bounds and malformed values.
• run.test.ts (fake clock): checks the rate-limited retry lines, the outage length and that no token reaches the terminal.
• e2e/build-restart.e2e.ts: the connector says how long Chopin was out of reach.