PROC-4 · Investigate cross-checking slowness in production

> Item derived from the fact-checking cycle analysis (**2026-07-03, v2**). ⚠️ Reconciled — read D1 below.

### Finding (reconciled — see D1)
Even in `main` alone, cross-checking went from **~6.6 h (2024) → ~69.8 h (2025–26)** — a **real** slowdown, not just composition. What *is* inflated is the "3× vs. 19.7 h" contrast: the 21.2 h historical figure is deflated by 2022–23 (~0 h, cross-check barely existed).

### Action
Investigate the cause (fewer cross-checkers available? more complex tasks? process change?) **before** setting a target.

### Reconciliation (D1) — real slowdown vs. cohort artifact (PROC-4 × INST-11)
Both, but separable. In pure `main`, 2024 (6.6 h) → 2025–26 (69.8 h) is **real slowdown** — not namespace composition. What *is* an artifact is the "21.2 h historical → ~70 h" contrast, because the historical blend is deflated by 2022–23 (~0 h). **Wording to adopt:** "there was real slowdown in production; do not cite the '3× vs. 19.7 h' as exact — always compare current state vs. 2024/2025 within `main`."

### Impact
Avoids optimizing the wrong stage; may reveal a capacity problem.

---
- **Priority:** 🟡 Medium
- **Dependencies:** a more precise reading depends on INST-2 (#2492); reconcile wording with INST-11.
- **Epic:** #2490