fix(web): narrow managed browser orphan reaping to provably orphaned processes

[cursor]

Follow-up from PR #1112 code review (open). The orphan reaper is the right direction, but the current startup path is too aggressive for production use.

Problem

On every daemon start, cleanupWebBrowserOrphans() SIGTERM/SIGKILLs all Chrome processes matching the managed install (launch marker or managed home path), with no check against:

• active daemon web sessions,
• agent-browser control socket / PID ownership, or
• recent browser activity.

After a quick daemon restart while a user still has a page open (or a long-running web workflow outside formal close), the new daemon kills the browser before accepting requests.

Additional gaps in the same PR:

• provider-startup cleanup reason is declared but never wired — orphans from a crashed web provider while the daemon stays alive are not reaped until the next full daemon restart.
• stopMatchedProcesses signals only top-level PIDs, not the Chrome process tree (renderer/GPU children can survive).
• Home-path matching may be over-broad (bare homeDir / runtimeHomeDir vs .agent-browser/browsers subtree).
• No tests for actual reap behavior (signal escalation, TERM→KILL, leave-unmatched-alone).

Suggested direction

• Reap only provably orphaned processes: no matching agent-browser socket/PID lock, idle beyond AGENT_BROWSER_IDLE_TIMEOUT_MS, or start time before last daemon shutdown marker. Skip reaping when session store reports an open web session for that stateDir.
• Wire provider-startup from the web provider factory on first resolve (with debounce).
• Walk process tree or kill process group for matched Chrome processes.
• Add unit tests with mocked ps / process.kill.

Acceptance

• Daemon restart does not kill a managed browser that an active session still owns.
• Crashed-provider orphans are reaped without requiring a full daemon restart.
• Tests cover reap signal order and process-tree coverage.

Context

• Open PR: fix: reap managed web browser orphans #1112
• Related issue: web: agent-browser Chrome fleets leak and can spin at 100% CPU for days — add idle lifecycle/reaper #1109 (if applicable)

[thymikee]

Resolved by the final merged #1112. The merged branch includes the narrow/hardened reaper commits (08263067, 0685d907): daemon startup passes open web sessions to the reaper, provider-startup cleanup is wired and failure-tolerant, socket activity can skip cleanup, process trees are signaled TERM to KILL, matching is constrained to managed browser markers/home, and the behavior is covered in agent-browser-lifecycle, agent-browser-provider, and daemon startup cleanup tests. Closing.