[TS Calls][Research] Capture stopped-path frontiers for article experiments

No files or tests are named. First identify a concrete consumer, research question, and fields unavailable from #364 decision events; if none exists, leave the work deferred. For an approved scope, define detached stopped-state ownership and bounded retention, then add focused tests for observer failure isolation, decision preservation, and separation from completed or replay-confirmed results.

Follow-up to #364. Part of #360. Optional research support for #385.

Goal

Capture stopped-state information only if an identified experiment requires data unavailable from existing unknown-call decision events.

Start condition

First name the consumer, the research question and the fields it cannot obtain by aggregating #364 events. Basic call counts, model decisions and stop-site frequencies do not justify a new snapshot API. If no concrete consumer needs a snapshot, leave this work deferred.

#364 provides decision callbacks only. It deliberately provides neither persistent path history nor stopped-frontier snapshots.

Scope

• Add opt-in capture orthogonal to the fallback setting only for the demonstrated consumer.
• Capture detached data rather than retaining live TsState/TsContext objects or allocation counters.
• Include the minimum stopping event, relevant signatures/statements and constraint status needed by that consumer.
• Add history or textual constraints only when explicitly required; do not serialize the entire engine state.
• Prefer experiment-owned aggregation/storage over changes to generic usvm-core primitives.
• Keep captured stopped states separate from completed execution states and generated tests.
• Define bounded retention and ownership with the consumer.

Definition of Done

• The consumer and required information are documented before implementation.
• Capture does not alter exploration decisions; observer failure isolation and detached ownership have focused tests.
• A stopped snapshot is never counted as a completed/replay-confirmed result.
• Deliver only the scoped research support in a separate PR when needed.

This issue does not block #365/#366, the census, or baseline evaluation in #385. It becomes a prerequisite only if the chosen measurement demonstrably requires it.