[Roxas2712]

Adds play_cry, text_sound and engage_music to API-3 portable scripts on generations 1–3. Cues are validated before admission, armed for the next show_text/ask, and played after typing through the native dialogue/audio controllers. Questions wait for audio before presenting their native choice; cries preserve auto-close versus explicit button-wait behavior.

Depends on #118 and its prerequisites. Merge that first. Focused review: https://github.com/Roxas2712/Gen2Recomped/compare/1fa9d59d90fbb6391a4a6cf994de56c016f40436...8a6eca8ebcc5501bc390f39de6aeba75b2878d48

Retained callbacks are once-only and inert beneath unrelated UI or after retirement. A reentrant unload during playback or source polling aborts the native update before it can pop a foreign overlay. Assets are revalidated at arm, construction and playback. Native 600-frame gates prevent stalled sources from trapping dialogue.

Validation:

• 112 targeted assertions covering all eight versions, malformed/missing cues, changed assets, audio/choice ordering, timeout, stale callbacks and reentrant sound/poll/music unloads.
• 160 cross-engine comparisons / 320 isolated runs against pinned Gen1Recomp; both clean feature and combined integration heads.
• Feature 93/103 engine suites; integration 98/108. Exactly the baseline 17 failed assertions and 10 failed suites remain; SDK 3/3 and B30 pass.
• Format/diff checks pass. Synthetic font/text data and a fake audio-device clock drive actual native dialogue controls; no imported-ROM or PCM equivalence claim.

Combined cry+text_sound plays both on the host. The reference Gen1 options merge drops the text sound in that combination; shared comparisons cover the individual cues instead of reproducing that defect.

Hosted workflows await maintainer approval for this fork (action_required): API 3, ci. No hosted pass is claimed.