[Roxas2712]

GBA Rage currently inherits the legacy rageMove controller, locking future choices and potentially exempting continued uses from PP. Its incoming-hit boost also runs inside generic HP damage, rather than at the native completed move boundary.

Implement a native rage flag: activate after accuracy, preserve native miss/immunity/cancellation/called semantics, spend PP on each separately selected use and leave the menu actionable. Raise Attack once for a completed qualifying opponent attack, before move-end item reads; honor Substitute, Endure, stage caps, ally/self/residual exclusions and a later immunity during multi-hit execution.

Clear Rage from actual effective choices before actions/turn-start observers, including Encore. Resolve double choices once and preserve native cached selections. Prepare one committed non-move reply for items, capture, switching and failed/successful escape cleanup. Add native AI STATUS2 visibility, stale-offer tracking and active link hashing; keep older-generation and explicit legacy controllers unchanged.

Stacks on #222 and reuses the existing #188 native AI/order prerequisite. A clean #222+#188 baseline is tested separately; #188 was already in the combined fork and is not reapplied there. Gen1 stays unchanged and read-only; no Gen1 development or publication.

Advertise GBA command revision 7, party codec 1. Independently audit raw fixture a7db0874db78aec0 → dcc938191b0bb075 using the unchanged #222 hasher with only the revision token changed. Older-generation goldens are unchanged.

Validation at clean feature runtime 03c4a841ee327d232ca504a1ae5df1c5159064b6 (later commit documents acceptance):

• 796 focused checks pass: 584 native command/planning/lifecycle, 122 host public-command/composition/reply/legacy, 90 actual link/compatibility. Native protocol and the reused AI/order suite also pass.
• 104 workers / 75,210 paired observations agree with clean Gen1 535a5ec088c1ae1bf4d91dd3e35a56f64cf46d91. The new mode adds 357 complete rows per FireRed/Emerald profile: all four actor slots, activation/failure/cancellation/called/custom moves, repeated PP, opponent/ally/self damage, multi-hit/Substitute/Endure/stage caps, later immunity, residuals and actual planners/Encore choices. Raw RNG bounds/values, accuracy-hook state, public move/damage events and all battler state observations remain retained.
• Move planning includes actual turn-word preparation with fixed choices. Non-move planning starts at the same explicitly prepared input boundary. Original unmatched planner reports are retained; this is not complete earlier/global selection/RNG/controller parity.
• Full feature: 123/132 engine suites, SDK 3/3, B30. The exact Counter of all 25 inherited FAIL lines across nine suites equals clean prerequisite fa5cb30ec4050980dc940e2181d6943a9de669b4 (120/129). The full command retains its nonzero exit.
• Combined runtime e1a9295e57c1545ec69deec2c4541a4da193b7b6: 243/243 engine suites, SDK 3/3, B30, the same complete comparison and four imported smoke scenarios pass.
• Unchanged prerequisite fails all four checks of the same public selector/delegated native continuation: missing native flag, legacy lock, only one PP spent across two uses, and locked menu. Raw negative/exploratory reports are retained.
• Two actual link roles complete four turns: Rage twice, a multi-hit versus retained Rage, then Tackle on both roles. All completed paired hashes/components, learned PP and stages agree; a whole incoming multi-hit builds Rage once and cleared owners get no stale boost. Revision 6 is refused before party exchange.
• Four CPU-only scenarios with actual local FireRed/Emerald imports pass real public FIGHT/move submission, repeated Rage (20→18 PP), alternative Tackle and the committed item-reply pipeline. Imported metadata is preserved with controlled valid accuracy rolls. No ROM/cache/save data is distributed or normal SAVE issued. The item scenario starts at the spent-item callback, not a full bag/healing playthrough.
• Scoped StyLua, actionlint and whitespace pass.

Host content overrides/delegation are tested separately. An extra synthetic Gen1 GBA catalog probe could not retrieve RAGE_EFFECT; its raw failure remains retained. This does not claim cross-engine effect-catalog/delegation equality. The native accuracy-stage adapter is private, not a new portable public effect stage.

See docs/native-gen3-rage.md for commands and boundaries. Broader canceller/full/custom/delayed scripts, Ruby/Sapphire scheduling, complete catalogs and item/capture/escape controllers, whole-game RNG/VBlank, rendering and full playthroughs remain separate work. The feature is pushed to Roxas2712/Gen2Recomped:codex/gen3-rage; combined publication and hosted CI are recorded separately from local acceptance.

Published combined snapshot: Roxas2712/Gen2Recomped:codex/portability-testing at 1d3d2c5311c31febe96c8679fb0b7cab238d9abb. Changes after the tested combined runtime only add validation/publication documentation. Fork main remains 13af468b97dc37015d14ad5c71d7f067c3ee810d.

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