Ship Group A (bug fixes) + Group B (pipeline enhancements) from
plans/260410-1009-openclaw-ts-feature-port — five changes that share
cmd/ wiring and pipeline plumbing so they commit as a unit.
cache: InMemoryCache now supports periodic sweep + max-size cap via
variadic options. PermissionCache wires 60s sweep + 10k entry cap +
Close() hook in gateway shutdown so long-running gateways don't leak
per-user permission entries. Backward compatible for zero-arg callers.
store: ContactCollector.seen cache key now includes tenant_id + channel_
instance so the same sender in different tenants (or different bot
instances in the same tenant) no longer silently skip upserts against
each other. Zero-tenant (Desktop) behaviour preserved.
pipeline: EffectiveContextWindow is resolved once per run in ContextStage
via a ResolveContextWindow callback (backed by providers.ModelRegistry)
so PruneStage bills history against the actual model window instead of
a stale static config. Nil resolver / unknown model fall back to
Config.ContextWindow for backward compatibility. Locked to the model
observed at context build time to prevent mid-run budget drift.
pipeline: PipelineConfig.ReserveTokens carves out an optional safety
buffer subtracted from the history budget so compaction fires slightly
before the hard limit — protects against provider over-delivery and
token counter drift on streaming responses. Zero (default) preserves
legacy budget math.
agent: ModelRegistry flows gateway → ResolverDeps → LoopConfig → Loop →
pipeline adapter so resolver can look up per-model capabilities at run
time without re-touching gateway internals.
20 regression tests across cache, contact collector, and pipeline cover
the critical paths: cross-tenant isolation, cache sweep + eviction +
Close idempotency, per-model window override + fallback, and reserve
token buffer behaviour. All passing with -race on both go build ./...
and go build -tags sqliteonly ./....
viettranx