feat: report paired benchmark differences

[cpunion]

Problem

A same-runner report currently compares the head median with the base median. That discards intentional sample pairing: runner drift can make the ratio of independent medians disagree with the median pairwise change. Reports also show only the percentage, not the signed difference in the metric's unit.

Changes

• add opt-in sample-pairing: index for benchmark files whose repetitions correspond by occurrence
• validate matching benchmarks, sample counts, and metrics at every paired sample index
• show every comparable metric as signed unit difference / percentage; paired reports use the median pairwise differences and percentage changes, for example -1.321 ns/op / -23.0% (better)
• keep the existing summary comparison when pairing is not declared
• preserve enough precision for very small nonzero differences and percentages
• index baseline benchmarks once per platform, keeping report rendering O(B·U)

Validation

• npm run check (33 tests)
• npm run build
• npm run benchmark
• rendered the LLGo #2027 Linux and macOS artifacts; an unchanged noisy benchmark changes from a misleading median-ratio +6.0% to a paired -0.11 ns/op / -0.9%, while the consistent interface-call result remains -1.321 ns/op / -23.0%

[fennoai]

FennoAI Review

Reviewed the sample-pairing: index feature (pairwise-median absolute + percentage change reporting for same-runner paired samples). The implementation is careful and well-tested: every rendering precondition is guarded before switching to the paired path, division-by-zero and undefined-percentage cases are handled, validateSamplePairing enforces matching benchmark/unit/sample counts on both write and load, and the dist/ bundles were regenerated. Security review found no injection or trust-boundary issues (units/keys are sanitized; baseline source is re-pinned to the PR target). Documentation matches the implemented behavior. All 19 affected tests pass locally.

One inline performance note below, plus a few minor observations. Nothing blocking.

Minor observations (no inline anchor needed):

• src/artifact.js validateSamplePairing verifies per-unit sample counts match but not that a unit occupies the same sample indices on both sides. In practice every Go benchmark sample carries all units, so this is unlikely to trigger — but the README claims the recorder "rejects ... rather than silently breaking the pairing," and a unit present at different positions with equal total count would pair mismatched repetitions. Consider asserting each sample carries the full unit set.
• src/presentation.js formatPreciseNumber: magnitudes below ~1e-7 render as 0 after trailing-zero stripping, so a tiny nonzero paired difference can display as +0 <unit> (sign from the original nonzero value). Edge display artifact only.
• The report note in src/report.js ("...pairwise medians") is keyed only on the current platform's samplePairing, while the actual paired path also requires the baseline to be index-paired with matching counts. validateSamplePairing keeps these consistent today, so the note is accurate; just coupled indirectly.

[github-actions]

setup-benchmark-go-action

8d26b1027220 | workflow run | long-term charts

Action bundle sizes

Bundle	Size	vs main
RecordBundle	330118 B	+1.1% (worse)
RenderBundle	330661 B	+1.1% (worse)

Parser and validation operations

Group	Benchmark	Median time	vs main
Grouping	GroupRules100	148480.834 ns/op	-35.3% (better)
Parsing	ParseGoOutput40	182503.771 ns/op	-30.0% (better)
Validation	ValidateArtifact40	198642.641 ns/op	-33.5% (better)

Compared only with the latest matching platform in the main series.

[cpunion]

All review observations are covered by the current head:

• c047c6a removes the repeated linear benchmark lookup, validates the unit set at each paired sample index, preserves tiny nonzero deltas, and only emits the pairwise-median note when both sides are index-paired.
• The inline performance thread is resolved.
• 8d26b10 keeps the wording and generated bundles consistent with the signed absolute/percentage format.

The latest CI and self-benchmark jobs pass.