[Roxas2712]

Each native palette zone currently allocates effective-color/uniform tables and sends all four uniforms; each zone also draws the full canvas through a scissor. P5 resolves native palette rows by index, compares twelve normalized floats in a per-shader private batch cache, reuses uniform vectors and crops supported zones with one cached Quad/setViewport.

Cache validity is limited to one contiguous native blit. Every later blit starts with a full upload; public sendColors calls always resend, and consumer-provided table send methods still receive fresh vectors. Existing public palette/graphics wrappers disable batching assumptions. Public effectiveColors/permute return behavior, c0–c3 uniform names, zone records and pipeline interfaces remain unchanged. Existing copyright/licence notices are retained.

Quad geometry is limited to identity transforms, nearest filtering, positive integer framebuffer magnification/origins, matching target DPI and ordinary render targets. Unsupported layouts retain the original scissor path. LÖVE 11 logical truncation and LÖVE 12 physical rounding are matched explicitly; the latter uses the quad path only at target DPI 1/2/4. Initial fractional-zoom crops failed the GPU oracle and were restricted/corrected before qualification.

Locations: src/render/PaletteFX.lua:1238 (public send), :1253 (native batch), src/render/Renderer.lua:708 (quad eligibility), :737 (crop), :760 (blit), src/core/Performance.lua:53. Independent rollback: require("src.core.Performance").paletteZoneCache = false (default true).

Stacked on #291, which includes #290 and profiler prerequisite #272. P5-only commit: 28e158f.

Validation:

• Fork: 430/430 engine suites, 3/3 SDK suites, 253/253 gallery checks; 200 contract comparisons / 400 isolated loaders against pinned reference 535a5ec.
• Authenticated fixture phase: 13/14 entry points load; the remaining follower fixture reports the exact named missing sprite-module gap allowed at this intermediate phase. Full follower support remains open.
• Public adaptation: 69/80 engine suites, matching all 29 inherited failure lines across the same eleven failing suites, with no added failure. SDK 3/3.
• 491 focused checks cover all display modes, shade mapping, palette mutations, changed-row sends, raw shader writes between blits, public overrides/wrappers, fresh consumer vectors and crop mapping.
• Native GPU readbacks: 432/432 byte-identical cases on each of LÖVE 11.5/OpenGL and LÖVE 12/Metal. OpenGL exercises 48 quad/384 fallback cases; Metal 16 quad/416 fallback cases. Cases include Red/Gold, display/shade modes, overlapping/true-color zones, clipping, transforms, three target densities, fractional origin/zoom and source filtering/density.

Native timings: macOS arm64, 60 Hz/VSync 1, 96 equal-palette zones, 120 warm-up + 120 measured frames per switch phase. Public F11 zone blit: OpenGL 0.41 → 0.22 ms/frame; Metal 16.37 → 16.35. Raw draw means: 0.445 → 0.243 and 16.376 → 16.364 ms/frame. Process CPU: 0.968 → 0.705 and 0.759 → 0.735 ms/frame. Heap growth with GC stopped over each 120-frame window: approximately 4368 → 49 KiB; this includes host/profiler work and is not retained memory. The equal-palette batch issues four Shader:send calls instead of 384. The fork's OpenGL run was slower, so no general FPS/CPU gain is established. Metal queue submission can absorb GPU/VSync waits. Full-game, Android, dedicated ARM-handheld and long-session: not measured.

Reproduce using the guarded tools/check_frame_pacing.py --palette-zones host. Raw F11 logs, measured-source fingerprints, API/backend rounding sources, both-tree measurements and limitations are described in docs/palette-zone-cache.md.

Contribution permission and credit: Roxas2712 contribution licence. Existing source notices and third-party rights remain effective.

Final composed acceptance and native Emerald profile

The completed composed fork runtime 9532ee36 passes 493/493 engine suites, 3 SDK suites, 253 gallery checks, 200 pinned API-3 comparisons and all 17 unchanged external entrypoints in complete mode. The complete public performance stack in #323 (runtime a620d0c7) passes 179/179 suites, 3 SDK suites, 253 gallery checks, 48 contract comparisons and all 17 complete entrypoints. It reuses repair PRs #128–#134 and #294. These results supplement the older standalone/baseline snapshots above; the eleven inherited failures are resolved on that final stack.

Real imported Emerald / Pacifidlog Town, actual native Game/Overworld and ordinary DRAMATIC_SHAPE loading on Apple M2 ARM64, LÖVE 12 Metal (profile source 79ff56cb): FrameProfile renderer: zone blit = 0.0215 → 0.0227 ms/frame, flag off → on, 240 measured frames per state and two real 120-frame profile windows. This is the real GBA zone path. It does not exercise GB c0–c3 uniforms; their separate byte/pixel and upload-count tests remain authoritative. The small change is within this sample's scene variability.

The guarded full-game starter freezes committed sources and uses a fresh owned ROM import/profile. It explicitly admits scripted QA into the existing native pacing/idle-GC body in its frozen copy; the timing implementation is unchanged. P1 uses live cap 60 / VSync 1, other phases the normalized native cap 160 / VSync off. The additional OpenGL replay passes all 14 native phases, then is stopped by the shared 20 GiB reserve during the optional voxel replay; no complete OpenGL voxel result is claimed.

These are sequential wall/submission-time samples, including slower values, not GPU completion times or a causal universal FPS gain. See the programme ledger and all samples/source hashes. Physical Android/Switch/handheld timings: not measured; the measured Mac is ARM64.

Current upstream qualification

Merged current upstream 8d5b4387; standalone 82/82 engine suites and all SDK suites pass at c8936955. Native Gen4 fog and poison-step bands are preserved across the cached zone-blit path (150 focused assertions). This early standalone host still lacks later acceptance tooling/fixtures and retains its inherited gallery failure; complete five-gate compatibility is qualified on the composed P9 stack (#323).