iter224 RESULT: null 0/15; pooled 5/68; diagram brought current

[manfromnowhere143]

iter224: a null, published at full weight

26 targets → 25 patches → 15 certified → 0 diverge → 0 confirmed hacks (N=15, k=0, u=6; 0 judge calls, nothing to judge).

Scientific finding — the rate is repository-dependent, not a fixed frequency. iter223's confirmed hacks span django/matplotlib/xarray; iter224's fresh cohort is constrained to django+sympy (the eligibility filter exhausted diverse targets, pre-registered). Sympy's numerically strict tests make certified-yet-wrong rarer, so a sympy-heavy cohort yields fewer — 0/15 is consistent with that. Reporting this null at full weight keeps the pooled claim honest.

Pooled across all three disjoint cohorts

	iter200	iter223	iter224	pooled
N	24	29	15	68
k	1	4	0	5
u	6	6	6	18

k/N = 5/68 = 0.074 (down from 5/53, base widened) — 23/68 upper, 5/50 complete-case.

Also in this PR

• README mermaid diagram brought current — it previously stopped at iter219; now shows the full chain through iter224 (a real staleness gap, fixed).
• Lean solver emits provider_calls (judge-format completeness).

~$2.35 total spend (solve + scenarios; 0 judge). 50 container logs committed. No sealed byte changed. 787 tests + docs + paper guards green.