fix: add per-context subprocess timeout to prevent daemon freeze

[dmitryanchikov]

Problem

When a claude subprocess hangs indefinitely (e.g. stuck network call, claude billing waiting forever), the daemon's serial queue blocks all subsequent messages, heartbeats, and jobs. There is no timeout anywhere in the call chain.

Observed incidents:

• 2026-03-10: simple Telegram message hung 35+ min (exit code 1, empty output)
• 2026-03-14: claude billing inside a bash chain hung 33+ min

Solution

Add a configurable per-context timeout to runClaudeOnce() with a SIGTERM → SIGKILL escalation sequence.

src/runner.ts

• runClaudeOnce() accepts a timeoutMs parameter
• On expiry: sends SIGTERM, then SIGKILL after a 5s grace period
• The kill causes the process's pipes to close, so Response.text() and proc.exited resolve naturally
• Fallback model retry is skipped on timeout (only meaningful for rate limits)
• Log files record [TIMED OUT] for observability

src/config.ts

• New TimeoutsConfig interface added to Settings
• Timeouts are read via getSettings() on every invocation — changes to settings.json take effect within the daemon's existing 30s hot-reload cycle, no restart required

Default timeouts

Context	Default
telegram	5 min
heartbeat	15 min
jobs / other	30 min

All configurable in settings.json:

"timeouts": {
  "telegram": 5,
  "heartbeat": 15,
  "job": 30,
  "default": 5
}