[CaelmBleidd]

Long PBT comparisons need original inputs and outcomes after generation, but point observations retain at most 64 invocations and repeatedly serialize their whole history for byte accounting. Add an opt-in compact input export with the existing tagged snapshot, phase, admission and outcome types. It records no point events or execution trees, and does not create a session or restart generation.

Inputs are snapshotted before precondition/predicate mutation. The recorder accounts only each new record's bytes, reserves final admission/outcome and counter sizes, and preserves explicit count/byte omissions, unsupported snapshots and pending asynchronous outcomes. Publication copies terminal fields so a late callback cannot turn a pending receipt into a completed one. Source identities, count, byte bounds, arity and omission counters are validated at the actual Kotlin process boundary. The opt-in export is bounded by 100,000 records and 2 MiB, within the existing 4 MiB wire limit; coverage attaches through its existing separate collector and old observations retain their existing bound. A transport/serialization error remains an explicit backend failure.

The API adds optional inputExportRequest/inputExport fields to the existing run configuration/result. Defaults retain the ordinary generation path and publish no input export. Overflow is evidence of incomplete capture, even when the original property run succeeds. This enables a later driver follow-up to use a large declared run quota and accountable budget stop; it does not itself establish an equal-budget experimental result.

Stacked on #586 at 6e632337a9106fdfe84b29dc95c4356df864ed82. No path selector, README or generic spool framework is added.

Validation: 218 Node adapter tests passed, including late-publication immutability and wire validation. All seven new real FastCheckBackend tests passed: 300 records alongside coverage/old observations, a response above 1 MiB through the actual transport, UTF-16 and nested arrays, precondition/predicate mutation isolation, count/byte overflow, unsupported snapshots/pending async work, and disabled defaults. The 21 nearby existing backend/invocation tests also passed. Main/test Detekt for both PBT modules reports zero findings. The driver still requires explicit complete capture before primary class comparisons.