[TheGreatAxios]

Summary

• Compaction occupancy is 60% of the context window minus the reserved completion budget.
• A follow-up whose usage or live estimate plus that reservation meets the window compacts before the next infer.
• An inference error that names a context length compact-and-retries even when the category is not already context_overflow. The recovery cap stays at 2. A message that only says "exceeds the maximum" or "reduce the length" does not fold the transcript.
• A server sentence that names maximum context length of N pins N for that model for the rest of the process.

Verification

• bun test src/agent/compaction.test.ts src/provider/context-window-thresholds.test.ts covers the reservation, the follow-up preflight, the two-retry cap, the loose-phrase negative, and the server pin.
• pre-commit tsc --noEmit and bun build pass on this commit.

Fixes CL-10833

[linear-code]

CL-10833

[corbits-triage]

Triage: Pass

No changes needed from you. A maintainer will review this pull request.