perf(pool): trigger a coalesced refill after consuming a warm VM

[CMGS]

Tracked from #30, especially the refill timing analysis in #30 (comment).

Problem

ClaimWarm removes a VM from p.warm but does not trigger refill. The first refill attempt therefore waits for startup, SetPools, or the two-second ticker.

The measured warm 0 -> 6 recovery was 1638 ms. The underlying recovery is:

• 0–2000 ms waiting for ticker phase;
• approximately 200–400 ms of actual refill work.

refillOne already self-chains once refill starts, so the missing lever is the initial kick, not a larger semaphore at this pool depth. This does not reduce the current warm claim's latency; it improves the warm-hit ratio for traffic arriving shortly after a drain.

Relevant code:

• ClaimWarm
• two-second refill loop

Proposed scope

• Send a non-blocking, coalesced refill signal immediately after a warm VM is removed.
• Let the manager Run loop consume the signal and call refillOnce.
• Keep the two-second ticker as a safety net.
• Do not spawn one goroutine per claim and do not block the claim response.
• Do not change maxConcurrentRefills in this issue.

A capacity-one channel or equivalent atomic pending bit is sufficient.

Acceptance criteria

• A warm pop starts refill without waiting for the next ticker.
• Concurrent warm claims coalesce rather than creating refill storms.
• Existing draining checks and the refill concurrency cap remain effective.
• A randomized-ticker-phase 0 -> 6 benchmark, sampled at 10–20 ms resolution, reports recovery near the actual refill work term rather than the ticker interval.
• Warm claim latency does not regress measurably.

[CMGS]

Fixed on main (db1bfa3): ClaimWarm sends a non-blocking, 1-buffered (coalescing) kick consumed by Run's refill loop. Bench host evidence (same harness as the #30 ledger, FC small pool warm=6): warm 0→6 recovery 1638 ms → 207 ms (−87%), i.e. the ticker wait is gone and only the actual refill work remains. burst16 wall 242 ms, warm hit 0.3 ms — in the historical band, no regression.