v2.4.12

Fifth slice of the 2026-04-19 audit fix wave. Single-focus release
closing I23 — the CE cold-start issue that caused the recurring
latency-gate flake in CI.

Fixed

• Cross-encoder warmup no longer poisons the rolling p95 latency
  window. The first _ce_rerank_timed call includes model loading
  (typically 15–40s for bge-reranker-v2-m3). Before 2.4.12 that
  first sample went straight into _CE_LATENCY_SAMPLES_MS and stayed
  there for the full 64-call deque rotation — the strict latency
  fallback at the call site then silently skipped CE for the next
  minute+ of queries without a clear operator signal beyond the
  missing cross_encoder_applied trace. Now the first
  _CE_WARMUP_SAMPLES calls (default 1; env BRAINCTL_CE_WARMUP_SAMPLES
  to override) are tracked separately, excluded from the p95 window,
  and surfaced as cross_encoder_warmup_ms in _debug_skips so the
  cost is still observable. This is what was causing the recurring
  latency-gate flake in CI over the prior release chain.

Testing

1878 passed, 28 skipped, 2 xfailed locally. Two new regression
tests at tests/test_ce_warmup_burn_in.py lock the warmup-exclusion
behavior and the BRAINCTL_CE_WARMUP_SAMPLES=0 opt-out path.