feat(combat): inject optional allowed-set into clean_damage_types/combatant_defenses (0.26.0)

[sligara7]

What

Adds a keyword-only allowed: frozenset[str] = DAMAGE_TYPES parameter to clean_damage_types() and combatant_defenses() in combat/combat.py.

• Default unchanged — intersects with the SRD-13 DAMAGE_TYPES exactly as before (backward-compatible; all existing call sites compile + behave identically).
• A caller that governs its own damage-type vocabulary passes a wider set (SRD-13 ∪ custom) so registered types survive the filter at component assembly, the computed-stats read, and live combat damage application.
• dndwright stays domain-agnostic (it accepts a set; it doesn't know where allowed comes from). Members must be lower-cased (caller owns canonicalisation).
• Public API (__all__) unchanged.

Why

Downstream StoryFlow v1.1.0 adds a per-universe damage-type registry. The hardcoded SRD-13 frozenset was the single chokepoint that silently dropped any custom type — and it is applied by two consumers (the character-sheet compute path AND the live-combat combatant_defenses call). Dependency-injecting the allowed-set is the single-source-of-truth fix that avoids a split-brain (registry honored on the sheet but not in combat) and avoids duplicating the filter across consumers.

Tests

• New real-path tests in tests/test_defenses_and_serialization.py: widened-vocab survival, default-unchanged back-compat, per-channel threading, and a registered type actually halving damage in apply_damage (governed at the point damage is applied, not just on the sheet).
• Full suite: 1655 passed, ruff clean, api-contract green (__all__ unchanged).

Minor bump 0.25.2 → 0.26.0 + CHANGELOG. Release/PyPI publish (gh release create v0.26.0) is a separate gated step.