v1.116.0

Changed — cache-busting release (CACHE_VERSION 4 → 5)

Every benchmark-level and over_time history cache key changes in this
release: expect a one-time cache miss and a fresh history series per the
standing cache policy (no migration of old entries). Implements plans 09 and
14 (plans/09-result-cache-invalidation.md,
plans/14-history-schema-reconciliation.md).

• BenchCfg.hash_persistent composition (plan 09):
  • result_vars and const_vars now contribute as an unordered set —
    reordering either no longer resets caches or history (D1).
  • Every per-variable identity (input, result, const) now includes the
    variable's name — renaming a variable is a detected identity change
    instead of a silent history split or phantom-dimension broadcast (D2).
• over_time history survives result-variable changes (plan 14): the
  history cache is keyed without result vars
  (hash_persistent(True, include_result_vars=False)) and stores a superset
  record of every column ever measured. At load time, columns are reconciled
  per identity (name, class, units, meaning_version) — added columns are
  NaN-backfilled and birth-stamped, removed columns go dormant (retained,
  invisible, resumed on return), redefined columns are retired under a mangled
  name and restart — and consumers receive a projection onto exactly the
  current config's columns. The benchmark-level result cache stays strict.

Added

• meaning_version on ResultFloat/ResultBool — bump it when a metric
  keeps its name but changes meaning; that column's history and regression
  baseline restart cleanly while every other column continues.
• BenchRunCfg.on_history_reset ("warn" default / "error" /
  "ignore") — loss-y history events (full reset with a named diff and
  orphaned-event count via a per-benchmark last-seen index, column dormant,
  column retired, incompatible history discarded) are logged, raised
  (bencher.HistoryResetError, before any state is persisted), or suppressed.
  Pre-record (bare-dataset) history entries participate in the same lifecycle:
  a column they carry that leaves the config is reported dormant, and it
  resumes rather than restarts when it returns.
• BenchRunCfg.regression_min_history (default 1 = previous behavior) —
  regressions on a baseline younger than N points since the column's birth are
  reported and exported (young_baseline: true in result.json,
  RegressionReport.has_blocking_regressions) but never trigger
  regression_fail; history-free absolute checks still gate. Per-variable
  override via a min_history key in regression_overrides. Young rows are
  marked notify-only (†) in the rendered regression report, the scorecard
  shows them as uncolored trend cells, and the exported report JSON carries a
  top-level has_blocking_regressions.