Iteration 3 — the last blocker RULE 0 · NO GATE RUN · 0 DECIDED

blockperm_pvalue returns (1 + cnt)/(B + 1), and the identity ordering always counts — it reproduces the observed statistic exactly. A window of nb blocks admits only nb! orderings, so identity recurs about B/nb! times and

E[p] >= (1 + B/nb!) / (B + 1)

Requiring that floor to sit at or below α derives the condition — no constant chosen anywhere:

nb! >= B / (alpha*(B+1) - 1)     at B=99, alpha=0.05  ->  nb! >= 24.75  ->  nb >= 5

4! = 24 does not clear it, and that is not a rounding quibble. Measured on a perfect signal — |ic| = 1.000, as strong as an effect can be:

L=300  nb=2   p=0.590   undetected
L=200  nb=3   p=0.240   undetected
L=150  nb=4   p=0.090   undetected      <- the intuitive floor lands HERE, and is blind
L=120  nb=5   p=0.010   DETECTED
L=100  nb=6   p=0.010   DETECTED

The handoff note from iter 2 guessed nb ≥ 4. Deriving beat guessing by exactly one block, and that block is the difference between a gate that works and one that reports zero power for a reason unrelated to power.

Sub-floor grid points are now refused before the permutation budget is spent — the honest answer is “cannot be measured here”, not a zero bought with 99 draws per window — and the artifact carries grid_points_refused_for_block_orderings and smallest_measurable_n, because an N_min from a partly-unmeasurable grid is a different claim from one fully explored.

With a NaN anywhere, obs is NaN and NaN >= NaN is False, so cnt = 0 on every draw and the p-value collapses to its floor. Measured on a genuinely independent pair:

clean                p = 0.590      correctly does not fire
one NaN injected     p = 0.010      FIRES, and ic itself is NaN

The failure direction is toward false positives, reported beside an effect size of NaN, in an append-only ledger. run_c3_and_power masked its inputs; run_c2 and choose_block_length did not — and the harness manufactures NaNs of its own, in the Y_delta residual and in known_positive's division.

Guarded fail-closed at the single choke point every fire decision passes through. A side effect worth stating: this makes D20 surface loudly instead of silently while it waits its turn.