fix(auto-resume): wait out rate limits, stop on other agent errors

[pratikgajjar]

Problem

A turn that dies on a provider rate limit ends through agent_end like any successful turn, so ensurePendingResume schedules the resume on the 800 ms settle window and the loop walks straight back into the cooldown. Every retry fails the same way until AUTORESUME_TURN_LIMIT or the consecutive-failure override trips — the session is spent, and nothing tells the user why.

Same shape for non-transient failures (bad request, context overflow, network): the resume message replays a turn that cannot succeed.

In practice the 429 often arrives as a bare JSON error body with no Retry-After header, so the wait has to be read out of the message text.

Change

agent_end now looks at the final assistant message before scheduling:

• Rate limited — resume is scheduled with the wait parsed from the error text (try again in 17 minutes, resets in 2 hours, retry-after: 45s), plus a 1 min buffer. No duration in the message means a 30 min default. The user gets a notification with the ETA.
• Any other agent error — no resume; notify and let the user decide.
• Clean turn — unchanged.

schedulePendingResume takes an optional delay (default SETTLED_WINDOW_MS, so every existing call site behaves exactly as before), and reschedulePendingResume reuses the stored delay so a settle-window reschedule cannot shorten an active cooldown.

Notes

• New pure module extensions/pi-autoresearch/provider-errors.ts (~55 lines) — no I/O, no session state.
• 4 new tests in tests/provider-errors.test.mjs; full suite is 50 passing.
• Verified by loading the patched extension in pi.