A benchmark example may carry its own cutoff

[tarekziade]

A run-level --before puts one instant on every example in a set. That is
right for a smoke test and wrong for a set whose examples come from different
points in history: twenty examples drawn from twenty issues have twenty
cutoffs, and a flag holds exactly one.

Example gains before (ISO 8601, timezone-aware or refused — a cutoff with
no zone is a different instant depending on who runs the set) and exclude,
both defaulting to nothing, so every existing frozen set stays loadable and
every untimed run is unchanged.

IndexSystem.window() narrows an example's window against the run's. An
example may narrow the run and can never widen it — the same rule, and the
same argument, as the AsOf daemon pin: a bound the run was started with is
not something the data it reads may lift. Exclusions are the union, because
neither side is entitled to readmit the other's withheld thread.

The signal post-filter is now pinned per example rather than per run: it
re-derives each hit thread's signals as of a cutoff, so a run-level one would
size the leak against the wrong instant for every example carrying its own.

20 tests. Suite 1326 passed, 4 skipped, 2 xfailed on both dialects.

🤖 Generated with Claude Code