INST-5 · Auditable classifications & adoption (umbrella: 5a + 5b)

> **Umbrella** · split into two child issues (see task list). Kept open as a tracking umbrella. Item derived from the fact-checking cycle analysis (**2026-07-03, v2**).

### Finding (evidence)
Cross → final adoption (old 98.7%) is **unauditable** — `report.classification` equals the initial value (in the 29 initial≠cross cases, the "final" adopts the cross **0×**; 20× stays on the initial, 9 have no final). Robust inter-rater initial↔cross: **84.5% (N=187)**.

### Split
- [ ] #2507 — INST-5a · Persist the reviewer's **initial** classification as a durable field — 🔴 P (quick win)
- [ ] #2508 — INST-5b · Store all three classifications + `agreement.*` booleans — 🟡 M

### Reconciliation (D2) — inter-rater adoption metric
Two numbers measure adjacent things: **84.5% (N=187)** is overall initial↔cross concordance; the **44.4%** (cross adoption in divergent cases) is **not reproducible** with today's fields — with `report.classification`, the final adopts the cross **0×** in the 29 divergent cases. The 44.4% needs an independent `final` that **does not exist today** — which is exactly what INST-5a creates. **Acceptance:** do not use 44.4% as fact until 5a/5b ship; only 84.5% is citable today.

---
- **Epic:** #2489

This issue is a tracking umbrella, with implementation split into #2507 and #2508; start by reading those child issues and their acceptance criteria. The work is done when the initial, cross, and final classifications and the agreement booleans are stored as specified, and the 44.4% metric is not treated as citable before those changes ship.

Rédigé par le modèle d'indexation à partir du texte de l'issue.