perf(tla): memoize the interpreted-path action text splitters

[zoratu]

split_action_body_clauses and split_action_body_disjuncts (action_ir.rs) are pure functions of their input text (the v2 layout-aware parse is deterministic; v2_enabled() is fixed per process), but they're re-invoked on the same action/clause text on every next_states call. During full-space exploration — and especially trace reconstruction, which re-explores the whole graph — re-running the Pratt parse + string scans per visit dominated the hot path for fairness/reconstruction specs once #189 removed the compiled-IR recompilation (gdb showed split_action_body_clauses / split_top_level / matches_keyword_at).

Wrap each splitter as an _uncached inner fn behind a thread-local memoization keyed by the input string (returns a cheap Vec<String> clone instead of re-parsing). Thread-local, so no cross-worker contention.

Result

ACP_NB_WRONG_TLC completes in ~29s (was ~143s post-#189 — a ~5x speedup), same verdict (AC1 violated).

Validation

• scripts/diff_tlc.sh: 20/20.
• cargo test --release: 1486 passed / 0 failed.
• Pure memoization of deterministic functions — no behavior change (confirmed by the gate).

Follow-up to #189 (which cached the compiled action IR in the guard path); together they remove the two dominant per-next_states re-work costs (recompilation + re-splitting).