[TS PBT] Evaluate runtime-derived PBT hints for symbolic search

[CaelmBleidd]

Part of #345. Depends on #354, #356 and the shared contract #384.

Goal

Test whether a small amount of concrete runtime feedback improves symbolic property search enough to justify keeping it.

Scope

• Start with at most one or two observation kinds motivated by the frozen development corpus, such as numeric ranges or array lengths. Do not implement every conceivable type/shape observation before measuring benefit.
• Reuse existing per-property run artifacts and a small optional observation payload. A backend without observations is supported by skipping this experiment.
• Define the selected hint precisely: its measured source, translation and whether it changes ordering or temporarily restricts the input domain. It never becomes a declared property precondition.
• For temporarily restricted search, run a bounded hinted attempt, then restart from the full declared domain without hints if no confirmed violation is found.
• Reserve a positive hint-free budget in advance and include observation, startup, hinted search, fallback and replay in the same total budget.
• A fallback with the remaining budget does not guarantee the same discoveries as a full-budget hint-free run. Report this cost rather than claiming preservation of completeness.
• Keep the hinted attempt and fallback separately attributable.
• Compare with [TS PBT] Assemble the end-to-end PBT and USVM pipeline #354's sequential HYBRID on the same properties, backend, seeds and total budgets. Include one useful and one misleading observation fixture.

Definition of Done

• The selected hint mechanism and budget split are frozen before the comparison.
• A misleading hint fixture demonstrates that fallback removes the hint restrictions and can recover a counterexample; it is not a general discovery guarantee.
• Results report replay-confirmed findings, time to confirmation, overhead and fallback frequency. State/solver counters are optional when already available.
• Negative results and unsupported observations remain visible.
• A short decision retains the tested subset, rejects it, or proposes one bounded follow-up.
• No general invariant miner, object-shape framework, relation mining or dynamic engine-selection scheduler is introduced.

This experiment is optional for the baseline article. #357 may proceed without it, explicitly recording that runtime hints were not evaluated. It must not block baseline evaluation merely because an optional production feature has not been built.