EPS Dashboard

Goal

Upgrade the plan-conditions coverage check from literal-slug coverage to a ran-but-unreported check: a registered condition whose results are present in committed result JSONs but absent from the clean-result body should flag.

Why

This exact shape produced a blocker in TWO consecutive clean-result rounds at #823, both caught only by an LM critic, neither by any mechanical check.

Round 8: the plan-registered rung-1 capped-vs-pure-GCV sensitivity read (sens_estimator) had run — /primary/5000/sens_estimator was present in the committed ladder_ext_r2.json with 30 fold-cells and a clean null — and the body had ZERO hits for sens_estimator, capped-vs-pure, or pure-GCV.

Round 9: the companion ladder's plan-mandated identity+learned-bias and kNN retrieval reads had run for BOTH ladders (228/228 cell keys, 4/4 kNN), and the body's only mandatory-reads table was stream-prefix-headed with the companion ranges appearing nowhere. The binding reconciler upheld this as blocking, on the ground that plan v17 states a REPORTING duty ("Baselines (mandatory, every fitted map, BOTH ladders)"; the reads "reported alongside held-out R² at every (rung × arm)"), not merely a compute duty.

The existing check flags literal slug coverage and WARNs advisorily, which both rounds passed while the substantive omission stood.

Scope

For each plan-registered condition slug, if its results appear in the round's committed eval_results/ JSONs but the slug's subject matter appears nowhere in the body, flag it. Paraphrase coverage must not trip it: #823's 12 WARN-flagged slugs were verified one by one as genuine paraphrase coverage, so a naive slug-grep would produce 12 false positives on a body that is actually complete. That is the hard part of this task and the reason it is not a one-line change.

Acceptance

A fixture where a registered condition's results exist in a committed JSON and the body never mentions the measurement FLAGS.
A fixture where the body covers the condition in paraphrase does NOT flag.