Republish with decisive judge diagnoses and grounded narratives (20260708) (#115)

Points the app at dashboard-data-20260708: the failure audit re-judged
under the PR #114 contract (per-model diagnoses, hedging mechanically
rejected, reference treated as verified), with 477 CHIP/school-meal/SSI
reference narratives regenerated grounded in engine facts after the
mechanism-required validator found 74 June narratives fabricating
mechanisms (e.g. CHIP cells claiming income exceeded a threshold when
the person is Medicaid-eligible and CHIP excludes Medicaid-eligibles).

0 hedged of 5,351 row annotations, 0 reference-suspect flags. Scores
unchanged — modelStats byte-identical to 20260707c — so the dataset
stays v1.1 and the hero label keeps Snapshot 2026-07-07.

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Author: MaxGhenis