feat(performance): unify baseline sampling and isolate collection #950

Stage 2 of #944: introduce one main sampling owner and an isolated, bounded collector. Reuse observations across the existing incident journal, freeze hooks, legacy diagnostics, and new cached monitoring API; prevent automatic heap snapshots during baseline use. Add one disposable renderer probe owner, bounded transport with failure/restart handling, and packaged helper verification. This stage preserves the current UI while the product monitor is built in stage 3.

Acceptance: one sampler per process, no duplicated event-loop reset windows, no unbounded in-flight batches, no source identity spoofing, graceful unavailable collector and shutdown, and deterministic lifecycle/fault tests. Packaged smoke and measured helper footprint will record practical limits honestly. Keep existing crash evidence available and keep legacy verbose tracing separately gated.

Depends on #949 / #948; refs #944 and #767. Storage/history, product UI, and advanced profiling remain subsequent stages.