feat: dirty-flag rule skip, mode-aware pruning, batch fact setter, condition cache

[tbitcs]

Performance Optimizations

Add four performance optimizations to the arbiter engine:

1. Dirty-flag rule skip (CONFIG_ARBITER_DIRTY_SKIP)

• Track which facts changed via uint64_t dirty_mask in snapshot
• Compute per-rule dependency bitmask at compile time
• Skip rule evaluation when no dependent facts changed
• Emit ARBITER_MODEL_RULE_DEPS in generated header
• Default on for standard/full profiles, off for nano/micro

2. Mode-aware rule pruning

• Add required_mode field to ARBITER_rule_def
• Compiler detects mode_guard rules with mode equality checks
• Engine skips rules whose required_mode != current_mode
• UINT16_MAX sentinel means evaluate in any mode

3. Batch fact setter (ARBITER_set_facts)

• Validate context once, then loop setting each fact
• Returns ARBITER_OK or first error encountered
• More efficient than N individual ARBITER_set_i32 calls

4. Condition result cache (CONFIG_ARBITER_COND_CACHE)

• 8-entry cache in eval loop for repeated condition checks
• Keyed by (fact_id, op, value) tuple
• Reset each eval cycle
• Default on

Tests

All 141 tests pass, 1 skipped (golden vectors placeholder).

---

Conversation: https://app.warp.dev/conversation/cbbf5a9b-9163-44ac-8a72-ca91f2f49f74
Run: https://oz.warp.dev/runs/019e8993-e9ad-7e1b-b02b-d75ab0d24caf
This PR was generated with Oz.