Scale the cpu-diff regression gate floor with the workload

[jantimon]

Problem

cpu-diff --fail-on-regression is the JS-cost gate the docs recommend, but its noise floor was a fixed 0.5 ms on the net JS self-time. That floor false-reds byte-identical code, and it gets worse with --iterations: summed self-time grows while the floor stays absolute, so a larger (or higher-iteration) workload trips constantly on sampling bursts.

Measured (all --target node, gate lock held, one measurement at a time)

Run-to-run net JS-self delta on byte-identical code, and the fixed-0.5 ms false-red rate:

workload	fixed 0.5 ms false-red	abs net noise (p95)	rel net noise (p95)
~5 ms (18 samples)	2%	0.49 ms	10.8%
~44 ms	27%	1.54 ms	3.4%
~220 ms (--iterations 20)	40%	6.27 ms	2.8%

Absolute jitter grows with self-time; relative jitter shrinks (finer quantization). A single absolute floor cannot fit both ends. This reproduces the integrator report exactly (red ~1 run in 4 at ~10 ms, worse with iterations).

Fix

The gate trips only when the net clears max(--noise-floor ms, --noise-pct% of the baseline), default max(0.5 ms, 15%). The percentage term tracks the absolute jitter as it grows; the absolute term keeps a small or 0 ms baseline from gating on a fraction of a sample. 15% sits above the worst measured identical-pair relative jitter (~11% at 5 ms) with margin.

New flags --noise-floor <ms> / --noise-pct <n> widen (noisier host) or tighten (large stable workload) each term. CpuDiffResult JSON now carries noisePct and the effective gateFloorMs, so a consumer reproduces the exit code from JSON.

Verified after the fix (built CLI, real recordings, actual exit codes)

• Identical pairs, default gate: 66/66 green at ~5 ms, 45/45 at ~220 ms, 28/28 on a same-path base-vs-base (100% green).
• Real +30% regression on a 5 ms workload (same module path, content changed +30%, so no comparability refusal): 64/64 caught across an 8×8 base-vs-current cross-product. A single pair misses only when the baseline recording itself over-samples high (a baseline-quality artifact the committed-once baseline avoids), not the floor being loose.
• The resolving floor (both sides below ~2 ms = not evaluable) still takes precedence; comparability refusals unchanged.

Scope

• src/commands/cpudiff.ts: the two-term gate floor + flag plumbing.
• src/cli.ts: --noise-floor / --noise-pct (validated by the existing numeric policy).
• src/model/query.ts: CpuDiffResult gains noisePct + gateFloorMs.
• Docs: cpu-profiling.md noise-floor section rewritten present-tense; facts.md row updated + a scaling-floor row added; README gate paragraph.
• Tests: 6 new gate-floor unit tests; measurement-lane comment refreshed.

Gates: lint, format:check, build, unit (744), measurement (6) all green. knip is unaffected by this change (it flags only worktree-local devDependency resolution here).