feat: recover empty and stalled responses

[wuxiaoqiang12]

Summary

• recover provider turns that complete with finish: "unknown", zero output tokens, no text, and no explicit error
• abort requests that remain busy for 200 seconds without text, reasoning, tool activity, or output tokens, then send continue after idle
• delay and revalidate recovery on idle, reusing the existing throttle and consecutive retry limits
• preserve user-initiated abort behavior and suppress retries while waiting for question or permission responses, including legacy and v2 events
• cancel the stalled-request watchdog as soon as real assistant activity begins, so long streaming responses and tool executions have no total-duration limit
• reschedule throttled retries instead of silently dropping them and dispose pending timers cleanly
• document retryEmptyResponses and stalledRequestTimeoutMs, and commit rebuilt dist artifacts

Motivation

Some provider failures do not emit a retryable session.error. One observed failure records an assistant message with finish: "unknown", zero output tokens, and no text, then emits session.idle. Another leaves the session busy for an unusually long time without producing any assistant activity. Both cases stop useful progress until the user manually aborts and enters continue.

This change uses strict per-turn evidence to distinguish completed empty responses and stalled first responses from normal replies, tool calls, user aborts, questions, and permission prompts. The stalled watchdog defaults to 200 seconds and can be disabled with stalledRequestTimeoutMs: 0.

Verification

• bun test (21 tests passed)
• bun run check
• bun run build
• loaded the rebuilt local plugin with OpenCode 1.17.20 in an isolated server configuration and verified /ac and /auto-continue registration