PTU action economy: ActionBudget models Full and conversions as independent buckets #359

Rulebook conformance audit found a production legal-action gap in Java.

src/main/java/io/autoptu/core/rules/ActionBudget.java currently mirrors Python by treating every ActionType as an independent consumed bucket. ActionType.FULL is therefore independent from STANDARD and SHIFT, and the budget does not express Standard→Swift or Standard→Shift conversion.

AutobattlerActionSpace consumes this budget directly when generating legal Shift and Move choices, so this can affect authoritative AI/controller legality rather than only tests.

PTU 1.05/Kairos action economy requires:

• one Standard, one Shift and one Swift Action per turn;
• Full Action consumes Standard + Shift, not a separate independent resource;
• Standard may be surrendered for an additional Swift;
• Standard may be surrendered for an additional Shift;
• the Standard→Shift conversion cannot be used for Movement if the regular Shift Action was already used for Movement, though it may still pay for non-movement Shift effects;
• Free Actions remain uncapped numerically, while Trigger restrictions are handled separately;
• Priority/Interrupt consumption needs its own lifecycle semantics rather than a free independent bucket.

Classification: RULE_LOGIC / likely PYTHON_RULEBOOK_DIVERGENCE pending exact Python resource audit.

Required implementation contract:

1. Represent base Standard/Shift/Swift resources explicitly.
2. Represent whether the regular Shift was used for Movement.
3. Make Full availability/consumption atomic over Standard+Shift.
4. Make Standard→Swift and Standard→Shift legal conversions explicit.
5. Preserve extra actions as named rule grants rather than treating them as substitutes for every base resource by default.
6. Update AutobattlerActionSpace to request rule-aware capacity, including whether a Shift choice is Movement.
7. Freeze rulebook-derived tests independent of Python before differential parity.

This is high leverage because it affects action economy, legal-action generation, AI policy, Trainer/Pokémon action resolution, and later manual combat.