Metric defect found by B's first run (mean = inf): an exact hit made the new width 0 and the floor (ulp of the root, 5e-324 when the root is 0) overflowed. Fixed to XTOL - an exact hit ends the run, so it is worth the halvings still needed to reach xtol - identical for every candidate; A re-run with it.

A with the fixed metric: still PASSES A1 and A2 in both configurations - flat 0.26 -> 2.82 (proposed) and 0.35 -> 0.77 (reading 3), canonical 6.32 -> 8.49 and 9.00 -> 10.44 - but a correction to the first report: for reading 3 on the flat families step A (0.77) beats our step and NOT bisection (0.99); the first run said 1.16. B (outcome memory) FAILS its falsifier: flat 0.31 -> 0.97 and 0.58 -> 0.82 (B1 holds) but canonical 4.74 -> 2.49 and 5.52 -> 2.12 (B2 fails), bisecting on 66% and 53% of canonical states. Mechanism: the known trap - a good step can collapse the error without moving the bracket, so an outcome measured in bracket shrink condemns it; the same trap that sank the bracket-progress gate (4449) and the bracket guard (7260).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

diag_step_extrapolation.py

@@ -21,9 +21,16 @@ XTOL, MAXFEV, ACC = 1e-12, 200, 1e-6
 
 
 def halvings_of(f, lo, flo, hi, fhi, landing, r):
+    """Halvings of the bracket bought by one evaluation at `landing`.
+
+    FIXED 2026-09-21 after the first run of PREREG B printed `inf`: an exact hit makes
+    the new width 0, and the old floor (ulp of the root, 5e-324 when the root is 0)
+    overflowed W/floor.  An exact hit ends the run, so it is worth exactly the halvings
+    still needed to reach xtol: the floor is XTOL, the same for every candidate.  A was
+    re-run with this metric; its first numbers used the old floor."""
     W = hi - lo
     Wn = bracket_step(f, lo, flo, hi, fhi, landing)
-    return math.log2(W / max(Wn, math.ulp(r) if r != 0.0 else 5e-324))
+    return math.log2(W / max(Wn, XTOL)) if W > XTOL else 0.0
 
 
 def replay(f, a, b, r, cfg):