[TS PBT] Challenge context-scoped observed function summaries during symbolic search #403

Implementation child of #345 and a required stage-two bounded extension. Owns helper-summary integration with the completed #355 core; it is not a core completion blocker. Builds on #397/#398 and the existing call-policy implementation; do not reopen or duplicate the TS Calls experiment in #360/#385.

Delivery stage

This is a stage-two extension of #345. Complete and record the core assessment in #405 before accepting this extension's end-to-end integration and results. Literature, interface design and focused experiments may start earlier; the extension never blocks #355 or #405.

Reuse the completed core contracts. This issue owns any extension-specific changes to observation, binding, scheduling, replay, shrinking and feedback integration, with focused follow-up PRs; do not retroactively broaden #351/#352 or require already completed core issues to reopen. A negative or inconclusive pilot is recorded honestly and informs design; it does not silently cancel this full-roadmap obligation.

Publish a separately identifiable extension configuration and evaluate it against the same original oracle with the extension disabled. Its final held-out evaluation belongs to #357 and is reported separately from the development pilot.

Goal

Test whether property-relevant observations at internal call boundaries can improve interprocedural search without promoting sampled behavior to a trusted universal contract.

Scope

• Implement a bounded subset of deterministic pure helpers with supported scalar/collection-length relations. Record callee/source hash, call site, caller/property context, input support and observed result relation.
• Derive summaries from original-runtime observations through [TS PBT] Infer property-relevant relational hypotheses from PBT observations #397. Preserve alias/mutation/read dependencies by restricting unsupported cases explicitly; do not model side-effecting helpers as pure.
• Primary use: prioritize call contexts and generate counterexample goals to a summary at the real helper implementation via [TS PBT] Project observed relations and assertion goals into exact EtsIR bindings #398.
• Evaluate speculative summary substitution only as an isolated explicit mode. It must carry assumptions into every candidate, have original-runtime replay and a budgeted no-summary search path. It cannot prove absence of bugs or justify marking paths infeasible.
• Refine/invalidate summaries with replay counterexamples and source/context changes. Keep different call contexts separate; one context's evidence cannot constrain another without justification.
• Reuse existing TS call models/policies and record them in configuration. Hold these policies constant across PBT-feedback comparisons so gains are not attributed to unrelated semantic-model improvements.

Definition of Done

• At least one context-specific helper relation is observed, challenged in USVM and refined or concretely refuted.
• A fixture detects misuse across call sites or outside observed input support; side effects and stale revisions are explicit.
• Required challenge/prioritization mode works without trusting summaries as hard contracts.
• [TS PBT] Evaluate causal contributions of property-aware feedback and bounded extensions #357 measures whether context-scoped summaries add to entry-only relations, including cost and misleading-summary cases. Speculative substitution may be rejected by evidence, with the negative result retained.