v1.112.0

Added

• BenchRunCfg.regression_overrides — per-variable regression specs. Maps result-variable name → a {method: threshold} dict drawn from percentage / adaptive / delta / absolute, or a bare number as shorthand for {'absolute': value} (a hard directional limit: minimize = ceiling, maximize = floor).
  • A listed variable is checked by exactly the methods in its spec instead of the benchmark-wide regression_method, so a threshold can be loosened or tightened per variable; an explicit {} opts a variable out of detection. Unlisted variables keep the benchmark-wide method, and names matching no result variable are skipped, so one override map can be shared across benchmarks.
  • Multiple entries run as independent checks — {'percentage': 15.0, 'absolute': 1.0} tracks the trend and holds a hard floor. absolute checks need no history and fire from the very first recording; the other methods skip until history exists.
  • Adaptive overrides: the threshold is the MAD limit (the dual-band percent gate still comes from regression_percentage); while history is too sparse for MAD the check skips instead of falling back to a percentage check, so a listed variable is never judged by a knob outside its spec.
  • Malformed specs never crash a run: unknown method keys and non-finite/bool thresholds are dropped with a warning, and a spec left with no valid checks keeps the benchmark-wide method (a typo can't silently disable detection). Bare-number shorthand accepts numpy scalars.
  • Motivating use case: CI tracking of flat health metrics (a success rate that must hold 1.0, an orphan-process count that must hold 0) next to per-metric trend thresholds on the same benchmark. Breaches flow through the existing machinery unchanged (has_regressions, regression_fail/RegressionError, report markdown, JSON, render_png).
  • First mutable (param.Dict) field on BenchRunCfg: every BenchRunner copy point deepcopies and the default is None; recorded in the copy-guard test's REVIEWED_MUTABLE_FIELDS allowlist with nested-dict isolation tests.

Changed

• detect_absolute with OptDir.none now warns and returns None instead of recording a non-regressed "checked" result — a guard with no direction never ran, so it no longer appears to have passed.