Search strategies¶

Parallel k-section bisect¶

With k workers, a round chooses up to k evenly spaced interior probes. The leftmost bad probe becomes the new upper bound; the rightmost good probe to its left becomes the lower bound. A monotone range of n commits takes at most ceil(log_(k+1)(n)) rounds.

Skipped probes do not move a boundary. Subsequent rounds borrow nearby untested commits. If every remaining interior commit is skipped, the correct result is a candidate set rather than a fabricated first-bad SHA.

A good result to the right of a bad result is non-monotone. The default policy stops and suggests scan; leftmost continues heuristically, while retry re-evaluates the contradictory probes once.

Parallel scan¶

Scan maps the evaluator over an ordered range and reports every good <-> bad transition. --stride selects every Nth commit. --sample selects evenly spaced commits and prints a follow-up bisect command around the first observed transition.