test(bench): document CI variance data for No-op and 1-file rebuild thresholds

The NOISY_METRICS set docstring explained the 50% tolerance for fnDeps depth 1
with CI run citations, but left No-op rebuild and 1-file rebuild documented only
by cross-reference ("same pattern as…"). Add concrete CI run IDs and
baseline→measured numbers for both metrics across v3.11.0–3.12.0, drawn from
the existing KNOWN_REGRESSIONS evidence already in this file, so future
maintainers can verify the threshold is empirically grounded rather than
arbitrary.

Closes #1516

Author: carlos-alm

tests/benchmarks/regression-guard.test.ts

@@ -65,6 +65,37 @@ const NOISY_METRIC_THRESHOLD = BENCH_CANARY ? 1.0 : 0.5;
  * only when its baseline is consistently sub-30ms and CI variance has been
  * empirically shown to exceed 25%.
  *
+ * - `No-op rebuild`: native baselines across releases span 15–30ms
+ *   (3.11.0: 15ms, 3.11.2: 19–25ms, 3.12.0: 23–30ms depending on suite).
+ *   Shared-runner jitter of ±10ms routinely lands +60–109% on these numbers:
+ *     - 3.11.0 baseline 18ms → 29ms (+61%) — run 26426483639
+ *     - 3.11.2 baseline 25ms → 45ms (+80%) and 19ms → 37ms (+95%)
+ *       — run 26792023287 (docs-only PR #1282, zero hot-path changes)
+ *     - 3.12.0 baseline 30ms → 48ms (+60%) and 23ms → 48ms (+109%)
+ *       — run 27457266151 (PR #1487 adds warmup runs, no-op path unchanged)
+ *     - 3.12.0 baseline 23ms → 114ms (+396%) — run 27455727444 (canary,
+ *       PR #1468 enclosing-caller fix; no-op exits before that code runs)
+ *   MIN_ABSOLUTE_DELTA (10ms) filters truly trivial jitter (e.g. 15→19ms),
+ *   but a runner under load can easily produce +20ms on a no-op, which is
+ *   both above the floor and above the 25% threshold on a ~20ms baseline.
+ *   The 50% tolerance absorbs this pattern while still blocking a genuine
+ *   regression that adds real work to the no-op path (which would show up
+ *   consistently across runs, not just on loaded runners).
+ *
+ * - `1-file rebuild`: native baselines span 64–115ms across 3.11.0–3.12.0.
+ *   CI jitter of ±20–50ms translates to +25–155% on sub-100ms measurements:
+ *     - 3.11.0 baseline 64ms → 109ms (+70%) — run 26706695868 (concurrent
+ *       PRs on same runner measured 64→80ms (+25%), confirming runner noise)
+ *     - 3.11.2 baseline 83ms → 212ms (+155%) — run 26793082961 (PR #1278
+ *       TypeScript resolver PR; locally measures 86ms, within baseline noise)
+ *     - 3.12.0 baseline 86ms → 131ms (+52%) — publish gate run (incremental
+ *       suite's identical metric PASSED in the same run, confirming cold-start
+ *       bias in the build suite rather than a structural regression)
+ *   The 50% threshold catches a genuine regression that adds real work (e.g.
+ *   an O(n) scan on every 1-file rebuild would consistently land +100ms+)
+ *   while absorbing the runner-load spikes documented above. Tracked in #1440
+ *   (add warmup runs to the build-benchmark 1-file tier to reduce this spread).
+ *
  * - `fnDeps depth 1`: native baseline 28.7ms (v3.9.6). The fn_deps Rust
  *   implementation, fnDepsData JS wrapper, and DB schema/indexes are all
  *   byte-for-byte unchanged since v3.9.6 (verified by `git log v3.9.6..HEAD`