fix: render platforms without paired baselines

[cpunion]

A pull request may introduce a platform that its base cannot build yet. In that case, requiring every result artifact to contain a baseline either hides the new platform or makes publication fail.

This change keeps same-runner pairing strict within each platform while allowing a complete platform to have no baseline. Reports retain paired deltas for platforms with baselines and render metrics from newly supported platforms as new, with an explicit note.

Validation:

• npm run check (39 tests)
• npm run build
• rendered the real three-platform artifacts from cpunion/llgo run 32520873182: Linux/macOS retain paired deltas and Windows renders all current program and core benchmark values as new

[github-actions]

setup-benchmark-go-action

6d4d1e394e5b | workflow run | long-term charts

Action bundle sizes

Bundle	Size	vs main
RecordBundle	330342 B	+224 B / +0.1% (worse)
RenderBundle	330885 B	+224 B / +0.1% (worse)

Parser and validation operations

Group	Benchmark	Median time	vs main
Grouping	GroupRules100	160426.746 ns/op	-10624 ns/op / -6.2% (better)
Parsing	ParseGoOutput40	195814.776 ns/op	+2416 ns/op / +1.2% (worse)
Validation	ValidateArtifact40	207071.205 ns/op	+661.6 ns/op / +0.3% (worse)

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

[fennoai]

Review summary

This change cleanly relaxes the same-runner baseline rule from a global "baseline required for every result artifact" to a per-platform, all-or-nothing rule: a platform may omit its baseline entirely, but presence must be consistent across all shards of the same platform. The implementation is solid:

• src/artifact.js uses a platformHasBaseline map with a symmetric, order-independent consistency check (previous === hasBaseline), and the specific per-platform assert message matches the updated test.
• Any baseline that is present is still fully validated (schema, shardId, platform, and sample pairing), so the relaxation drops only a coverage/completeness guarantee, not an integrity one.
• baselines.length === 0 ? [] : mergeShards(baselines) correctly guards mergeShards against an empty set.
• Untrusted fields rendered in the report note (baseline.source.sha, baseline.source.url) remain validated (40-hex SHA, HTTPS-only URL wrapped in <...> autolink) — no injection surface introduced.
• Test coverage is good: rejects inconsistent shards within a platform, allows a whole platform without a baseline, and verifies the new marking plus the note text.

One documentation follow-up below. No blocking issues.

---

Findings

[P2] Stale README: baseline behavior no longer matches the per-platform logic

README.md (lines 424–430) still documents the old all-or-nothing model:

When every platform artifact includes baseline-benchmark-file, each metric is compared with that paired baseline measured by the same runner job... Otherwise, each metric is compared with the newest matching platform in main.

After this PR that wording is inaccurate. With mixed platforms (some with a same-runner baseline, some without), the paired platforms render vs base while the unpaired platforms are marked new (metricComparison returns an undefined baseline) — they do not fall back to the main series. The comparison decision is now per platform, not a single global switch. Consider updating this paragraph to describe the per-platform behavior and note that platforms lacking a paired baseline are marked new, matching the new report note in src/report.js.