decision.altitude_floor: measure what the name says, and drop a prior the archive does not support

[pokle]

Branch preview: https://claude-altitude-floor-metric.glidecomp.pages.dev

The metric was labelled "How high the pilot commits to the next climb" (Commit%) and explained as the height still in hand "at the moment they commit to a climb" — but the computation never looked at a climb. It swept the 30 s-smoothed altitude trace for local minima with 100 m of prominence either side.

It is now "How low the pilot goes before the next climb" (Floor%), computed by walking each consecutive PAIR of post-start climbs and taking the lowest fix between them, keeping the gaps that descended at least 100 m. Same question, structural answer: "between climbs" is the loop bound rather than something the prominence sweep happened to produce, so the run out to the first climb, the glide to goal and a sled run are excluded because nothing was climbed after them. No smoothing, no prominence, and the reported low is a real fix altitude rather than a 30 s mean that flattens a deep save upward.

Audited over the whole archive

184 tasks, 25 comps, 4,762 pilot-tasks from pokle/glidecomp-archive, scoring each task and running the real metric.

• The two definitions agree and separate equally. Spearman 0.87 on pilot ordering; median 40.6% of band against 38.2%; median |ρ| against GAP rank 0.43 against 0.40, stronger on 51% of tasks — a coin flip. This change is made for explainability, not for signal.
• Coverage rises to 85.0% of started pilots from 83.7%, but shifts which pilots qualify: three climbs are now needed for two gaps, where two dips used to do.
• The direction: 'higher' prior does not hold, so it drops to neutral (the v9 treatment). Only 69% of 147 task correlations take the assumed sign, 16 are wrong-signed at |ρ| ≥ 0.3, and two large fields are emphatically opposite — Bright 2023 open T3 (85 pilots, ρ +0.73, top ten at 33% of band against the bottom ten at 78%) and Bright 2024 open T2 (95 pilots, ρ +0.52). Whether height in reserve pays is a property of the day, so the sign is the finding.

Two things the audit cleared rather than changed, both recorded in comments so the next reader doesn't re-ask:

• The window runs to landing, not ESS. Only 8% of ESS finishers have any post-ESS dip, and capping at ESS moved the median by more than 5 points for 2% of them. A gap needs a climb on both sides, so the glide to goal is excluded by construction.
• The old definition was not measuring wobbles inside climbs. Only 0.1% of 39,272 minima sat more than 60 s into a thermal; the rest were at a climb's boundary either side.

The explanation also now states the band's real endpoints — p10 of the field's thermal entries, p90 of its exits — so a negative reading reads as expected rather than alarming. Over the archive 2.3% of scored pilots sit below 0%.

Notes for review

• FIELD_ANALYSIS_VERSION 18 → 19 rolls stored task_field_analysis reports onto the new label, direction and values.
• Explanation text is written in the Simplified Technical English house style this file adopted in Simplified Technical English across the field analysis, and metric labels that name the quantity #519, which this branch merges.
• decision.altitude_floor moves to the 55% coverage tier in the kosci integration gate. That is a fixture property, not a scope one: the synthetic tracks are climb-poor (15 of 41 started pilots take fewer than three climbs, against an archive median near nineteen). Real-archive coverage went up, not down.
• New test pins the distinguishing behaviour: a pilot who climbs three times and then glides to the deck reads at their between-climb lows, not at the altitude they landed from.
• bun run test:all green (engine 1229, frontend 518, auth-api 92, competition-api 477) and typecheck:all clean.

🤖 Generated with Claude Code

[github-actions]

Preview Deployment
https://a8563cb3.glidecomp.pages.dev
Commit: 5ea51ef