Spend imminent overflow mana safely

[adrunkhuman]

Summary

• Forecast the earliest actual mana-regeneration cycle across food, equipment, and other regeneration conditions, including finite final ticks and same-cycle aggregation.
• Outside hunting, cast one explicitly audited safe spell only when that cycle would lose mana above the maximum; exact full mana and protection-zone pauses do not trigger.
• Prefer useful Haste, Great Light, then Light, with cheapest audited refresh-safe fallback, while preserving 20 emergency mana and normal engine legality.
• Arbitrate the opportunity below service, rewards, spell learning, and equipment but above the next hunt, then immediately resume normal goals without a training window or persistent state.

Verification

• pwsh -File scripts/test-playerbot-gameplay.ps1 -MagicTraining -Focused -SkipBuild - passed 16 forecast, spell preference, overflow, arbitration, failure, continuation, and restart scenarios
• pwsh -File scripts/test-playerbot-gameplay.ps1 -SpellUse -SpellCalibration -AdaptiveChallenge -Focused -SkipBuild - passed existing spell-use, calibration, and hunt-pressure regressions
• pwsh -File scripts/test-playerbot-gameplay.ps1 -FullNavigation -CorpseLoot -Focused -SkipBuild - passed navigation and corpse-loot regressions
• pwsh -File scripts/test-knight-spell-contract.ps1 - passed audited Knight spell contract

Limits

Magic-training scenarios use controlled Lua setup; they prove engine forecast arithmetic, arbitration, and selected normal casts, not long-running live frequency. Regeneration phase intentionally resets through condition serialization, so restart recomputes a fresh forecast.

The existing -SpellTraining route fixture passed once on this branch but timed out on repeated runs after its spell-learning transaction failed and normal equipment/service behavior continued. Failure logs contained no selected magic-training candidate, so this appears unrelated to #98.

Closes #98.