Steer exploration toward assertion and numeric boundary states

[chaoz23]

Product outcome

Turn author-defined invariants such as gold >= 0, health <= max_health, and “required flag before ending” into safe deterministic boundary probes while continuing to evaluate assertions on every visited state.

Parent: #97. Related: #23/#64 assertions, #65 goals, #87 causal distance.

Bounded behavior

• Identify variables/operators/literals from the existing typed assertion grammar.
• Prefer observed transitions approaching, crossing, or resetting a boundary.
• Treat equality, off-by-one, min/max, boolean prerequisites, and terminal scope separately.
• Preserve baseline exploration; directed work is explicit/additive until Build the search-strategy promotion benchmark harness #56 proves a safe allocation.
• Never execute arbitrary expressions or infer domain meaning from names.

Acceptance criteria

• Numeric probes cover observed values immediately below/at/above supported boundaries when reachable.
• Boolean/terminal/knot-scoped rules receive matching specialist tests.
• Exact counterexamples include observed values and indexed replay paths.
• A clean bounded run still says no violation observed, not verified.
• Fixtures cover decrement loops, resets, max relationships, unreachable boundaries, non-monotonic updates, and violations behind sparse prefixes.
• Unknown variables/types/unsupported arithmetic fail before spending specialist budget.
• Matched evaluation reports target gains and all unrelated runtime/assertion/outcome/knot losses separately.
• No zero-sum default steering before Build the search-strategy promotion benchmark harness #56 passes; additive mode remains available for controlled evidence.

[chaoz23]

Authored-story evidence for #111: a root-started assertion child on The Intercept found the credible gotcomponent=true && piecereturned=true lead first at state 190,773, while its 100K base did not evaluate the rule. The same child ran to 1,472,319 states / memory and also found one semantically weak rule.

The steering evaluation should therefore report time/state-to-first validated violation, not raw violation count. It should include current-state variables versus historical-event booleans, preserve an unsteered assertion baseline, and require exact witness identity parity so prioritization cannot lose a later critical rule while chasing an easy boundary.