[Roxas2712]

Beat Up currently selects only the active trainer's party. In a two-trainer or player/ally battle, native Gen1Recomp execution reads the entire side party: six eligible members contribute six hits even when the user belongs to the second trainer.

Read primary slots followed by partner slots through a private live view. Keep the owning arrays and replacement benches separate. Honor the battle-scoped player party, current user's live HP/status, and replacement of an upcoming bench member inside a public hit callback. Capture the current array references for the move, matching the reference when whole party fields are reassigned mid-move.

Stacked on #209. Review c2a1a03 relative to 03d68a4.

Validation:

• 1,728/1,728 focused checks: 18 cases at all four positions on FireRed/Emerald; six/five members, status/egg/faint filtering, either owner or the whole side blocked, accuracy-time eligibility, callback waking/blocking/replacement, whole-array reassignment, live/detached active battlers, scoped player view, Substitute, fainting, criticals and variation.
• The shared API-3 mod and fixture use actual native constructors: Gen1 State with ownership halves; Gen2 trainer constructor, setAlly, addOpponentTrainer and sendOutSecondPair with prebuilt imported-shaped members. Ownership guards and replacement benches are verified separately from Beat Up's side read.
• The complete unfiltered comparison against unchanged read-only Gen1Recomp 535a5ec088c1ae1bf4d91dd3e35a56f64cf46d91 passes 48 isolated workers / 4,766 paired observations. Actual actions, PP, hit order/amounts, event metadata, callback counts and counter memory are compared.
• Negative control: clean Verify native Beat Up held-item adjustment across engines #209 fails 58 assertions in the new FireRed cases.
• Full feature quick run: 91/100 engine suites against 90/99 on clean Verify native Beat Up held-item adjustment across engines #209. Exact Counters of all 25 inherited FAIL lines across nine suites match; no additions. SDK 3/3 and B30 pass.
• BEAT_UP_EFFECT revision 2 intentionally changes the fixture digest from d518411055f17444 to fdbfa30b20cbbaef. Only move_effects changes (f45a80ce5e45ef42 to e4d28f474c15cd8b); restoring revision 1 restores the former digest.
• Pinned JIT-enabled LuaJIT a4c064539ea931274d407e1249b4bb575424a0f4, scoped StyLua, actionlint and git diff --check pass.

These are native actions and constructors with synthetic imported-shaped parties, not a graphical playthrough or a full link/multi-owner controller claim. A scoped player view and detached active battler are seeded explicitly; no new selection/replacement API is introduced. Full raw RNG traces, arbitrary mid-move ownership reshaping, intros/animations, other scripted executors and complete gameplay remain separate acceptance surfaces. Gen1 stays read-only, with no source edit or publication.

Combined full-suite acceptance and hosted CI are reported below.

Hosted CI: all six workflows for c2a1a03 report action_required pending maintainer approval. The dedicated contract run has zero executed jobs. Local acceptance is separate.

Combined-fork acceptance: integrated into Roxas2712:codex/portability-testing. Clean runtime f39cc19 passes 217/217 engine suites, 3/3 SDK suites and B30 with zero FAIL lines. The same clean runtime passes all 48 workers / 4,766 paired observations. Integration updates only the audited move_effects section pin; the fingerprint gate passes 33/33 checks, including the unchanged sections and mutation checks. Subsequent combined commits only record validation.