fix: prevent excessive background LLM requests causing rate limiting and sluggishness

[BYK]

Problem

Users reported multiple symptoms pointing to excessive background LLM requests:

1. Heavy rate limiting from upstream LLM providers
2. Slower LLM interactions
3. Many "error" sounds in TUI with no visible indication
4. "Prompt too long" errors briefly visible for very large numbers
5. Overall sluggish OpenCode server behavior

Root Causes & Fixes

Bug 1: Auto-Recovery Infinite Loop (CRITICAL)

When a context overflow (prompt too long) triggered auto-recovery via session.prompt(), if the recovery itself also overflowed, a new session.error fired with no re-entrancy guard, creating an infinite loop:

overflow → distill + recovery prompt → overflow → distill + recovery prompt → ...

Each cycle fired 2+ LLM calls, repeating until rate-limited.

Fix: Add recoveringSessions Set. If a session is already recovering when a second overflow arrives, bail out immediately (forceMinLayer is still persisted for the user's next manual message).

Bug 2: Curator Fires on Every session.idle (HIGH)

The condition used onIdle || turnsSinceCuration >= afterTurns. Since onIdle defaults to true, the || short-circuits — afterTurns (default: 10) was never evaluated. The curator fired an LLM worker request after every single agent turn.

Fix: Change || to && — curate on idle only when enough turns have accumulated.

Bug 3: shouldSkip Lists All Sessions on Every Unknown Message (MEDIUM)

When session.get() failed (common with short IDs from message events), the fallback called session.list() fetching ALL sessions on every unknown message event.

Fix: Remove session.list() fallback, cache sessions as known-good after first check. Worker sessions are already caught by isWorkerSession(). Accept the tradeoff: short-ID child sessions won't be skipped, but a few extra temporal messages from eval are harmless.

Testing

• 14 new tests in test/index.test.ts covering:
  • isContextOverflow detection for all error message formats
  • buildRecoveryMessage with and without summaries
  • Re-entrancy guard: concurrent overflow events for the same session (only 1 recovery prompt fired)
  • Curator gating: no curation when turnsSinceCuration < afterTurns
  • shouldSkip: no session.list() fallback, sessions cached after first check
• All 187 tests pass (173 existing + 14 new)