fix: correct mask writer mixing and instance-hierarchy mask composition

[MasamiYitsuse]

Two masking defects, both confirmed against SS7.5 by SSPlayerForUnity first and ported here. Companion to SpriteStudio-SDK#341, which fixes the guide sample this player had been written from (already merged; the submodule bump is included).

1. Invert and increment writers shared one accumulator

Invert and IncrementWrap are each their own inverse in isolation, but they do not commute. The CBP replay folded both into a single 8-bit stencil, so wherever writers of differing mask_influence overlapped, coverage depended on the order the writers were emitted in — increment then invert landed on 254 (covered) where invert then increment wrapped back to 0 (not covered).

Split into a parity accumulator (mask_influence == 1) and a counter (mask_influence == 0), unioned at the end. Two inverted masks cancel, three cover (parity — not "exactly one"), and any increment writer covers outright.

Unlike a GPU stencil, the replay divides no bit budget, so ForUnity's "7 overlapping masks" ceiling (a consequence of URP's reserved bits) does not apply here.

2. Instance settings replaced the callee's instead of composing

SpriteStudio composes mask settings down an instance hierarchy: visible_inside_mask with OR, mask_influence with AND. The player decided both from the instance part's own flags and stamped one polarity across every material the child owned, so the sub-animation's parts never took part. It disagreed with SS7.5 in both directions — a callee opting out with mask_influence == 0 was masked anyway, and a callee's visible_inside_mask was flattened to the caller's.

The composed context is now pushed down before the child builds, the child bakes the resulting polarity into its own per-part materials, and the coverage owner binds only coverage / meta / UV / rank. The identity (influence = true, visible_inside = false) leaves a top-level player resolving to its parts' own flags, so unnested playback is bit-identical to before.

Godot-specific wrinkle: whether the mask reaches an instance is known only after the coverage is rasterized, which happens after the children build. That gate stays where it was; beforehand it is predicted from static data (does the part tree hold any writer). Over-predicting costs a per-part material that is then disabled; under-predicting would drop the mask outright.

This also uncovered a gap the blanket stamp had been hiding: the three emit paths gated the decision on this player owning coverage, so a sub-animation with no mask parts of its own could not be an inherited target at all. They now accept an inherited context too.

Verification

Visually checked in the editor against SS7.5 using tests/overall/Mask.ssae: mixed mask_influence overlap now unions, same-influence pairs still cancel, and a callee's own settings survive composition. Both builds green (GDExtension + custom module).

Known limitations (unchanged / documented)

• Multi-level instance nesting is unverified against SS7.5 — as it is in the SDK guide (2-6) and in ForUnity. The implementation chains by passing the composed value down a level at a time.
• A child that owns masks and is inherited-masked still resolves to the inherited one. Pre-existing (one coverage sampler), and treated as a separate matter by the guide.

Notes

The submodule bump carries documentation only — no code, schema or header changes, so no rebuild is required for it.