Gap-spacing prediction — 2/5, with the diagnosed fix never built

Explicitly speculative, and filed mainly because the diagnosis is untested. A prediction that scored 2 of 5 with a stated cause is worth more than one that scored 5 of 5 with none — but only if the stated cause is then checked, and it has not been.

The diagnosis. The search performed was unconstrained; the correct search is boundary-slack-constrained. On that reading the three misses are cases where the unconstrained search reached configurations the constrained one would never have proposed, so the rule was being scored against the wrong candidate set rather than being wrong.

The friction. The constrained version was never built. So the diagnosis is untested, and there are three readings with nothing between them:

1. The diagnosis is right — the constrained search recovers 5/5, or close.
2. The diagnosis is post-hoc — it explains the misses because it was constructed to, and the constrained search does no better.
3. The constraint helps but does not close — the rule is partly right, and the residual is the actual content.

What would settle it. Build the boundary-slack-constrained search and rescore the same five cases. Seal the predicted score first — predict the number out of five before running it, because a diagnosis that was going to explain any outcome is not a diagnosis. That is the whole test, and it is cheap.

Why this is a probe and not a run. The search is not specified precisely enough to launch: OPS.md's bar at draft needs a question, a prediction with a confidence, a falsifier, and what it settles, and "boundary-slack-constrained" is not yet a definition. Specify it and this becomes a run.