[TS PBT] Capture bounded property-scoped observations in original TypeScript #396

Implementation child of #355. Builds on #395, #348/#349/#384. Design and basic probes can start before #354, #356 or #382 complete.

Goal

Capture actual program values and their property context from concrete PBT runs, beyond aggregate coverage and a final boolean.

Pilot scope

The first delivery supports #405: numeric/scalar and dense bounded-array values, arguments/returns and selected explicit intermediate points, bounded snapshots and original assertion provenance. Restrict automatic instrumentation to the selected supported syntax/points and record annotation cost; whole-program automatic instrumentation is not a pilot prerequisite. Unsupported values and side effects are diagnosed instead of sampled unsafely. Dedicated cross-execution pairing, command-state contexts and helper-summary inference are accepted in #401/#402/#403, respectively.

Scope

• Extend the existing per-property artifact with run/input identity, provenance, property/assertion ID, generator identity, source/build hash, precondition admission, result/exception, phase (generation, explicit example, symbolic replay, shrinking), and available coverage.
• Observe selected function arguments and returns, pre/post values, assertion operands, and selected intermediate values at stable program points. Begin with function boundaries and explicit observation points; add supported automatic instrumentation using the existing TypeScript toolchain. Report manual annotation cost.
• Pair pre/post snapshots by invocation and distinguish repeated call sites, nested calls and property context. Reuse tagged values and supported alias representation. Arbitrary graphs/getters/proxies are not silently JSON-coerced.
• Instrumentation must evaluate expressions once and preserve this-binding, evaluation order, short-circuiting, throws, supported mutation isolation and async classification. Do not invoke getters or user serialization merely to record a value; mark unsupported observations.
• Bound points, collection elements, trace size and retained examples with deterministic policies. Record truncated/dropped observations and capture overhead; absence is not zero or a null value.
• Retain admitted/rejected/failing provenance without accidentally learning a program invariant only because a precondition or the tested assertion filtered the data.
• Mark correlated shrink/replay samples and solver-generated samples. They may refine hypotheses but cannot inflate the count of independent random evidence.
• Export common artifacts suitable for [TS PBT] Infer property-relevant relational hypotheses from PBT observations #397 and [TS PBT] Project observed relations and assertion goals into exact EtsIR bindings #398; no new process framework.

Definition of Done

• Real fast-check runs capture at least one input-to-result and one intermediate relation relevant to an original user property.
• Instrumented/uninstrumented executions have matching supported outcomes on fixtures covering mutation, special values, exceptions and short-circuiting.
• Isolation across properties, phases and invocations is tested. Stable provenance permits source-to-EtsIR binding and rejects stale builds.
• Bounded retention, unsupported values, instrumentation failure and missing coverage produce explicit outcomes.
• Capture works with no branch mapping and no completed campaign orchestrator; full-run overhead is available to [TS PBT] Evaluate causal contributions of property-aware feedback and bounded extensions #357.