[Roxas2712]

Triple Kick currently repeats one accuracy/critical/damage calculation three times at twice its extracted power. Native GBA effect 104 checks every reached kick, assigns 10/20/30 power and calculates damage independently; rounding, misses and move-end state make the approximation observably different.

Implement the native per-hit script through a per-dataset builtin adapter. Stop at accuracy/protection/hidden/immunity, sleep/faint and actual Endure boundaries; continue through Substitute, retain Focus Band and spend one learned PP per selected attack. Expose each assigned power through the existing public damage hook's ctx.opts.power. Preserve explicit damage/count/full overrides, delegation, neverMiss, shared statics and older-generation behavior. Keep completed Rage behavior when a later kick misses, and suppress it only when immunity is actually reached.

Stacks on #225. Gen1 is an unchanged read-only reference; this PR neither requires nor publishes Gen1 changes. Native command revision 8 rejects revision 7; party codec stays 1.

Validation at clean feature runtime 10a5c325a7852c6e8f7fea1502fafd9cfdc4a393 (later commit only documents results):

• 2,220 focused checks pass, plus the 154-check native protocol suite.
• 108 workers / 76,362 paired observations agree with clean Gen1 535a5ec088c1ae1bf4d91dd3e35a56f64cf46d91; the new mode adds 576 complete rows per FireRed/Emerald profile, all actor slots, each possible missed kick, critical stages/items, public per-kick power overrides, live lifecycle conditions and Rage. Raw RNG bounds/values, hook observations, damage/move events and battler state are retained.
• Full feature: 126/135 engine suites, SDK 3/3, B30. The exact Counter of 25 inherited FAIL lines across nine suites equals freshly tested clean Implement native GBA Rage state and move-end boosts #225 (123/132); the full command retains its nonzero exit.
• Combined runtime c1dac0994aa9fe0d1659e510874065697dca2534: 246/246 engine suites, SDK 3/3, B30 and the same complete comparison. Independent baseline repairs are in this combined branch.
• Unchanged Implement native GBA Rage state and move-end boosts #225 fails 240 assertions using the exact final fixture; 396/576 complete FireRed rows differ. Exploratory/negative failures remain retained.
• Two actual link roles complete three turns through real menus with every paired completed hash/component and learned PP equal. Revision 7 is refused before party selection. Raw fixture hash dcc938191b0bb075 → b9ec4e8a94f4ba6e is independently audited with the unchanged Implement native GBA Rage state and move-end boosts #225 hasher and only the explicit revision token changed; other-generation goldens stay unchanged.
• 12 CPU-only scenarios with actual FireRed/Emerald imports pass public move submission and completed turns: all three kicks, misses on 1/2/3, Substitute break and an actual faster enemy Endure. Extracted attack metadata is unchanged. No ROM/cache/save assets are distributed and no normal SAVE is issued.
• Scoped StyLua, actionlint and whitespace checks pass.

The earlier missing RAGE_EFFECT probe is resolved as a catalog boundary: pinned Gen1 seeds no GBA native move-effect records; Gen2 exposes its builtins. The same minimal authored primary record registers in both public loaders. Host native builtin delegation is tested separately and is not claimed as a shared cross-engine catalog contract.

See docs/native-gen3-triple-kick.md for reproducible commands, accepted/negative evidence and limits. Full Sleep Talk/other controllers, complete catalogs, Ruby/Sapphire per-hit scheduling, global RNG/selection orchestration, rendering and full playthrough parity remain separate work. Hosted CI execution and combined publication are recorded separately from local acceptance.

Published combined snapshot: Roxas2712/Gen2Recomped:codex/portability-testing at 5bbe0dbdfff2fe664c1a2f451ce0be51204f952e. Changes after tested runtime c1dac0994aa9fe0d1659e510874065697dca2534 only document acceptance/publication. Fork main stays 13af468b97dc37015d14ad5c71d7f067c3ee810d.

All nine hosted workflows await maintainer approval (action_required). The dedicated contract run has zero jobs, separately verified; local acceptance does not claim hosted execution. The app's existing 100-artifact identity cap rejected an attachment, but this regular PR is published and linked.