Show a family's behaviours as one chart, not just a table

[pokle]

Stage 1 of #458 — the per-family chart. The cross-family chart builder follows as a second entry point into the same component; see the proposal.

Preview: https://claude-glidecomp-issue-458-v.glidecomp.pages.dev

Merged up to master (ca777d2) on 29 Jul — clean, no conflicts, and everything re-verified against it. See "Testing" below.

The problem

"The metrics in detail" gave each family a wide per-pilot table and some prose. Charts appeared only where the engine emitted an extraSeries, and only 2 of the 26 metrics do — so four of the six families had no chart at all, and nothing answered the question a family is: was this pilot good across the whole of climbing, or only one part of it?

The table's columns are in different units (pct, m/s, s, km/h, ratio) and cannot be compared with each other. The heatmap compares everything at once but as luminance, with no numbers — and since #485 it no longer groups by family at all, so there was no per-family reading left on the page.

The chart

MetricLanes draws one lane per behaviour on one shared 0–100 percentile axis, one dot per pilot, and joins one pilot's dots down the lanes.

TopOut% †  ····•·····•·········  29/32     ← Jon Durand's climbing profile:
  -16%          61%       94%              topped out high, kept most climbs,
Kept%   †  ······•···········     28/32     found cores fast, left lift early,
Core s  ↓  ········•·······       32/32     circled less roundly than the field
LeaveRate †  ··········•···       32/32
Round   ↓  ····•···············   31/32
Out-climb ↑  ······•···········   32/32
           0    25   50   75  100
              percentile in this field

Exactly one connector line ever exists, which is why this is lanes and not parallel coordinates: forty polylines over six axes is spaghetti, and the usual fix is to stop drawing thirty-five of them. It breaks across a lane where the pilot has no value rather than inventing one — null here is "not applicable", never a score of zero. Hovering anyone (a dot, a table row, a heatmap row, a cluster chip — PilotHighlightContext is already page-wide) redraws it for them.

Two decisions that came from driving it, not from the plan

• The lanes default to the winner's profile. A percentile axis spreads the field near-evenly by construction, so with nobody hovered the chart was rows of dots and little else — which is all that print, a touch user, and a reader who never hovers would ever get. Defaulting is the same "strongest finding on first paint" that SeparationRanking and LegWaterfall already do. Print gets a line naming whose profile is drawn, since the hover readout cannot exist on paper.
• Lanes run strongest-separator first, reusing the orientedRho() that Order the field-at-a-glance columns by what separated the field #485 added, and outcome-derived metrics get no lane — theirs would run cleanly best-to-worst and read as a finding about flying. Both match the reordered heatmap. The per-pilot table below keeps registry order, so the caption says what orders the lanes.

What the caption has to say out loud

• A percentile axis says nothing about spread. Hence the value labels at each lane's ends (the real scale, in the reader's units) and deliberately no quartile band — on this axis it is always 25–75, a mark encoding zero information.
• Right is the end a behaviour is expected to be better at, never a claim that it paid on this task — that is the ranking's question. RankScatter.captionText already refuses the prior when a day runs the other way, and the two must not appear to contradict.
• Neutral metrics are not oriented and wear †, as in the heatmap. A lane's count is the pilots it applied to.

Percentiles are midrank, matching every other percentile on the page — the heatmap cell, the ⓘ popover strip, the cluster medians. The engine's private percentileColumns() uses (rank−1)/(n−1) and stays private; mixing the two would have a lane disagree with the heatmap for the same pilot on the same page.

Testing

All numbers below are post-merge, against ca777d2.

• All filtering, ordering and prose lives in a pure metric-lanes.ts with 24 tests, following PercentileHeatmap.test.ts. Covered: the ordering and its determinism under reversed input, no mutation of the caller's array, the outcome and blank-lane filters, lower flipping while neutral does not, nulls dropped rather than zeroed, perPilot in a different order from report.pilots still pairing by trackFile, zero-variance and single-pilot lanes, and every caption clause appearing only when it applies.
• The sort's id tie-break is a plain codepoint compare, not localeCompare — these pages are SSR'd and collation differing between workerd's ICU and the browser's is a hydration mismatch.
• bun run test:all — 1228 engine/script, 540 frontend vitest, 476 competition-api, 92 auth. bun run typecheck:all clean.
• bun run test:e2e:ssr — 32 passed, including the :taskAnalysis clean-hydration check that the deterministic lane sort guards.
• Driven in a real browser against the seeded Corryong Cup 2026 field, before and after the merge: 5 lane figures render (the day family correctly gets none — it has one per-pilot metric), the connector zig-zags across all 6 climbing lanes, the readout reads correctly, and the page-wide highlight reaches the table below. The rendering is pixel-identical pre- and post-merge, so Field analysis: put the ranking table and its chart side by side #489's side-by-side ranking layout doesn't disturb it.

Accessibility

Follows the heatmap's contract: one role="img" figure whose accessible name reads every lane in words (min/median/max, coverage, the †, and any missing top-3 pilot via the existing notableExcludedRanks) and names the table below as the data equivalent. The dots are deliberately not focusable — six lanes of forty pilots is 240 tab stops for values that are already in a keyboard-navigable table a few pixels down. Hit targets are 18px for WCAG 2.5.8; the connector is --foreground ink rather than a chart hue, since the dots carry identity and the line is an annotation.

🤖 Generated with Claude Code

https://claude.ai/code/session_01R1Zrf2tMEP2QHK6jWboAQq

[pokle]

Merged up to master (2cb994c) and verified the lane charts against glidecomp-archive: 184 tasks, 1,104 families, 3,239 lanes.

New web/scripts/audit-metric-lanes.ts, following audit-score-charts.ts. It scores every task in a comp library with the real engine, builds the real field-analysis report, runs buildLanes over each family the way MetricFamilySection does, and checks the invariants the drawing code relies on:

• dots pair by trackFile, never by index
• every percentile sits strictly inside (0, 100) — the midrank property the caption tells the reader about
• the value printed at a lane's end is the value of the dot at that end (the one most likely to be silently wrong, since a lower-is-better lane is flipped)
• the order is deterministic under reversed input (an SSR hydration requirement)
• no outcome metric gets a lane

Zero invariant violations across the archive. The three fixes below all came from things the audit counted, not things it failed on — each renders as a blank or an average-looking dot, which is exactly why no unit test on synthetic metrics was going to catch them.

1. A lane measuring one pilot is dropped — 136 of 3,401 lanes

Midrank puts a lone value at exactly 50, so the lane stated "this pilot was dead average" when it meant "only one pilot could be measured". There is no field for a percentile to sit within. The coverage count disclosed the n, but position is what a reader reads first, and at n = 1 it is false. The value is still in the per-pilot table, which is where a single reading belongs.

Side effect: it halved the all-identical lanes, 283 → 127.

2. The default profile falls past a leader with nothing to draw — 5 of 861 charts

It was rank 1 unconditionally. A winner who never shared a thermal has no gaggle metrics at all, so those charts drew no connector — precisely the rows-of-dots state the default profile exists to prevent. It is now the best-ranked pilot appearing in at least two lanes, since a single dot is not a profile and the connector needs two points to exist. Zero such charts remain; 5 charts now legitimately draw someone other than the winner.

3. An all-identical lane says so

127 lanes stack every dot on P50 under a bare figure, which reads as a broken chart rather than as the finding it is. The label now reads "every pilot 0.4 m/s". The metrics that produce them are unsurprising — decision.low_saves (38), race.start_delay (37), the gaggle pair (44 between them).

Also adopted from master

• Field analysis readability: the plan, and phase 1 — title every chart's axes #501 — the shared XAxisTitle/YAxisTitle from charts/AxisTitle.tsx, since every chart now names its axes in the plot ("a caption is read once; an axis is consulted continuously"). Verified in a browser that both clear the label gutter.
• decision.altitude_floor: measure what the name says, and drop a prior the archive does not support #520 — decision.altitude_floor went higher → neutral. No change needed: the orientation is data-driven, so it simply stops being oriented and picks up the †.

Verification

bun run test:all (1279 engine/script, 584 frontend, 508 competition-api, 92 auth), bun run typecheck:all, bun run test:e2e:ssr — 33 passed. Re-driven in a real browser against the seeded Corryong Cup 2026 field.

Run it yourself with:

bun --tsconfig-override ./tsconfig.json web/scripts/audit-metric-lanes.ts
GLIDECOMP_COMPS_DIR=<archive>/comps bun --tsconfig-override ./tsconfig.json web/scripts/audit-metric-lanes.ts

(The --tsconfig-override is needed because the script imports frontend chart code, whose @/ alias bun otherwise won't resolve from web/scripts/.)

One thing I did not change, for the record: 891 lanes still measure fewer than 8 pilots — the threshold below which the engine refuses to claim a correlation at all. At n = 2 or 3 a percentile is coarse but not false (it still says who was ahead), and each lane prints its own coverage count. Dropping them would have removed 27% of all lanes and gutted whole charts, so the count discloses it instead.

---

Generated by Claude Code

[github-actions]

Preview Deployment
https://71aa82e1.glidecomp.pages.dev
Commit: d3294a0