pi rehouses its documentation-lift evals in fresh containers and exports extension event hook types

Both pull requests pi merged on September 15, 2026 come from christianklotz, and both shore up infrastructure rather than the agent loop. #9635 rebuilds the documentation-lift evals around fresh Docker containers, and #9642 exports the extension event hook types that issue #9511 found missing — plus one Bedrock cache-cost regression fix riding along. Neither is in a tagged release yet; both will ship with pi’s next one.

#9635 separates host evals from the Docker documentation-lift comparisons and runs every planned (case, variant, model, repetition) arm in its own fresh container, reporting blocked pairs instead of aborting the whole cohort. The suites themselves are simplified around native describeEval, explicit run calls, deterministic judges, and live ModelRuntime checks, and eval scenarios and fixtures now sit next to the eval definitions that use them.

#9642 closes #9511, which catalogued the patchy state of the type exports: BeforeProviderRequestEvent was importable while AfterProviderResponseEvent was not, and the issue tabulated the gaps event by event — ResourcesDiscoverEvent, SessionBeforeSwitchResult, MessageEndEventResult, ModelSelectEvent, ThinkingLevelSelectEvent, among others. The fix exports every extension event and result type used by ExtensionAPI.on() from the package entry point, adds MessageEndEventResult and ThinkingLevelSelectEvent to the extensions barrel, and ships ModelSelectSource alongside ModelSelectEvent. Some types remain internal; the PR notes it may be worth formalising the internal/exported boundary later.

Riding along in #9642, and flagged as unrelated by its own summary, the Bedrock cache-cost regression fix makes that check respect profile-specific catalog pricing regenerated during CI. The full missing-type table sits in #9511 and the container-per-arm plan in #9635 — both are short reads before the next release cuts.