ci(perf): cache the main baseline binary to cut the nightly double build (#4816)

Every push to main now builds the release binary once and stores it in the
actions cache as rustfs-baseline-<sha> (build-baseline-cache job). The A/B job
restores it for origin/main and passes --baseline-bin; on the nightly, where
the candidate commit equals the baseline, one cached binary serves both phases
with --skip-build and the run does zero source builds. A cache miss falls back
to the source double-build via --baseline-ref origin/main.

This removes the ~32min-per-side double build that pushed the 24-cell nightly
past its 120min ceiling (2026-07-11..07-14 all cancelled on timeout).

Also:
- alert-on-failure now fires on cancelled/timed-out, not just failure, so a
  timed-out nightly is no longer silent (the composite action already reports
  cancelled jobs); removed the stale perf-2 TODO now that the alert job exists.
- run_hotpath_warp_ab.sh appends a Provenance section to gate.md (baseline and
  candidate SHAs + binary source, runner, warp version, matrix params) via a
  repeatable --provenance-note; the gate exit code is preserved.

Refs rustfs/backlog#1152 (perf-3).

docs/operations/hotpath-warp-ab-runbook.md

@@ -37,15 +37,31 @@ Six workloads × the drive-sync on/off matrix × baseline/candidate = 24 cells:
 Sizes are passed to the load driver via `--sizes` (one size per cell); the
 driver's `DEFAULT_SIZES` covers 1KiB..10MiB, so any of those can be added by
 editing `WORKLOADS` in `scripts/run_hotpath_warp_ab.sh`. A 1KiB cell is left
-out for now to keep the nightly matrix under the 90-minute job budget (the
-baseline+candidate double build dominates until perf-3 caches it); re-enable it
-once that lands.
+out for now to keep the nightly matrix comfortably under budget; re-enable it
+(one line in `WORKLOADS`) once perf-6 recalibrates the warp params.
 
 CI runs a **short** warp matrix (`--duration`/`--rounds`/`--cooldown` tuned in
 `.github/workflows/performance-ab.yml`) so all 24 cells fit the budget without
 dropping cells. These params are deliberately noisy-but-fast for the Phase-0
 "keep the pipeline alive" goal; perf-6 recalibrates them.
 
+## Baseline binary cache (CI)
+
+The nightly no longer builds both binaries from source. Every push to `main`
+runs a `build-baseline-cache` job that builds the release binary once and stores
+it in the actions cache under `rustfs-baseline-<sha>` (perf-3). The A/B job
+restores the binary for `origin/main` by that key and passes it as
+`--baseline-bin`; on the nightly, where the candidate commit equals the baseline
+commit, the same cached binary serves both phases (`--skip-build`) and the run
+does zero source builds — the common path finishes well under 50 minutes. A
+cache miss (binary evicted, or not built for that SHA yet) transparently falls
+back to the source double-build via `--baseline-ref origin/main`.
+
+Each `gate.md` ends with a **Provenance** section recording the baseline and
+candidate commit SHAs and whether each binary came from the cache or a source
+build, plus the runner, warp version, and matrix params. `perf-5`/`perf-12`
+reuse this contract for their archived baselines.
+
 ## Local mode (quick / CI smoke)
 
 Builds both binaries and runs a throwaway single-node server on local disks.