fix(viewer): compare-mode bridge KPIs render both arms in section + Notebook

[thinkingfish]

Summary

Three related bugs along the compare-mode rendering path for category-bridge KPIs (e.g. `inference-library` TTFT/Prompt Token Rate, where the dashboard generates a vllm-side query and a sglang-side query for one chart). Hit at the same time when pinning a bridge chart into the Notebook.

1. Asymmetric label extraction for percentile (scatter) charts

Baseline read `spec.series_names` (string list, lossy on multi-dim metrics — only the first non-`name` label survives) then ran `canonicalQuantileLabel` over the string. Experiment read `item.metric` directly via `canonicalQuantileLabel(item)` which preferred `metric.quantile` and dropped extra dims. The two sides produced disjoint match keys → "compare: no shared labels between captures."

Fix: `composeScatterLabel(metric, options)` in `compare_math.js` — multi-dim aware, with an `excludeValues` set (the category bridge) that drops dims whose values are capture-identity names. `applyResultToPlot` now also stashes `spec.series_metrics` so the baseline path has the raw metric objects available. Both extracts forward `options` symmetrically. Falls back to `canonicalQuantileLabel` for malformed metrics.

2. Pin loses bridge metadata

Pinning a bridge chart kept only `spec.promql_query`. Re-rendering in the Notebook ran the baseline query on the experiment side too, returning empty.

Fix: entries gain two optional fields, `promql_query_experiment` and `category_members`, captured at pin time from the spec + Group's `sectionMetadata`. Round-tripped through localStorage persist/restore and JSON export/import, defaulting to null on older payloads. `chartLoaderMixin` restores `promql_query_experiment` onto the rebuilt spec; NotebookView passes `category_members` to `CompareChartWrapper` as `categoryMembers`.

3. Pinned charts node-inject the wrong host

NotebookView mounted `CompareChartWrapper` with `sectionRoute: '/notebook'`. `buildEffectiveQuery` skips node-label injection for `/service/*` routes (service KPIs are scoped by `source=...`, not node), but `/notebook` doesn't qualify, so the experiment-side fetch ended up with the baseline's hostname injected into the bridge query — zero matches against the experiment recording, server returned "Metric not found". Only the baseline trace rendered.

Fix: use `entry.section` (the chart's home route) instead of `/notebook`. Bridge KPIs now skip node injection symmetrically with their original section view. Non-bridge pinned charts also keep their original section context — slightly different from before but more consistent.

Plumbing

`renderServiceSection` forwards section-level metadata as `sectionMetadata: meta` (a new prop, not `metadata` which is reserved for per-group config). Group reads `attrs.sectionMetadata?.category_members`, threads it through chartBody → CompareChartWrapper → both extract* functions → composeScatterLabel.

Test plan

• cargo build clean, viewer smoke 9/9
• node tests 47/48 (the 1 failure is the pre-existing wasm comparison test)
• 15 new unit tests in compare_math.test.mjs covering composeScatterLabel: pure quantile, quantile + N extra dims (alpha-sorted), no-quantile fallback, malformed input, excludeValues option (drop matched, preserve non-matched, empty/missing → no-op), baseline=experiment symmetry assertion
• Manual: `view vllm=vllm.parquet sglang=sglang.parquet --category inference-library` — TTFT and other bridge KPIs render both arms in the inference-library section
• Manual: pin a bridge KPI from the section, navigate to Notebook — both arms render (was: only baseline; experiment query had `node=""` injected and returned empty)

Known limitations

• The `composeScatterLabel` excludeValues filter is by VALUE (drops any dim whose value matches a category-member name). Footgun: if some unrelated label happens to carry a value like `vllm` or `sglang` (e.g. `binary=vllm`), it gets dropped too. In practice this is fine — the category-member names are distinct enough.
• Multi-dim percentile compare for non-bridge sections (e.g. per-cgroup percentile) is no longer broken (was the original sub-bug 1) but isn't smoke-tested end-to-end — the unit tests cover the helper and the section/group plumbing pattern is straightforward.

🤖 Generated with Claude Code