fix(gmicloud): alias prompt to text/lyrics for GMI audio models

[jeronimodeleon]

Summary

Every GMI Cloud audio model (TTS and music) via genblaze-gmicloud's GMICloudAudioProvider was unreachable — every request 400'd with "Required parameter is missing" because the SDK's own param allowlist stripped the API's required text (TTS) / lyrics (music) fields before the request went out, and there was no alias from the canonical prompt param (the one every other modality uses) onto them.

Changes

• libs/connectors/gmicloud/genblaze_gmicloud/models/audio.py: added a _TTS ParamSurface (_AUDIO_BASE.with_aliases(prompt="text")) used by the gmi-audio-tts family, and chained .with_aliases(prompt="lyrics") onto _MUSIC for the gmi-audio-music family. with_aliases() auto-adds the alias target to the allowlist, so text/lyrics are no longer stripped. Updated the module docstring to match.
• libs/connectors/gmicloud/tests/test_catalog_decoupling.py: registry-level regression tests confirming text/lyrics are allow-listed and aliased from prompt for the TTS/music families.
• libs/connectors/gmicloud/tests/test_gmicloud_audio_provider.py: end-to-end regression tests confirming prompt= reaches prepare_payload() and the outgoing HTTP body as text (TTS) / lyrics (music), with prompt itself absent from the final payload.
• docs/reference/model-matrix.md: updated the affected rows' "Allowed params" column (this file is generator-owned by tools/gen_model_matrix.py; hand-patched here to match exactly what the generator would emit, since this sandbox lacks the API-key env vars the generator needs to enumerate every connector without wiping unrelated sections).
• CHANGELOG.md: [Unreleased] entry under ### genblaze-gmicloud.

Deliberately out of scope: the gmi-audio-clone (MiniMax Voice-Clone) family was left untouched. The issue's own root-cause analysis and suggested fix are scoped to gmi-audio-tts/gmi-audio-music only (its title literally says "the TTS/music param allowlists"); voice-clone was mentioned in the issue's affected-models table without a reproduced failure. Rather than guess at an unconfirmed payload contract, I filed a scoped follow-up: #257.

Test plan

• make test passes (verified this session — full suite, 19 packages, 0 failures)
• make lint passes for the touched connector (ruff check/ruff format --check libs/connectors/gmicloud/)
• Docs updated in this PR (CHANGELOG + model-matrix.md)
• make typecheck — not run this session (scoped to libs/core, unaffected by this change; connector typecheck is optional per the release-check gate)

New regression tests (test_catalog_decoupling.py, test_gmicloud_audio_provider.py) were confirmed to fail against the pre-fix code (KeyError: 'text'/'lyrics') and pass after the fix, per TDD.

This PR went through a triangulated 3-lens review (correctness/tests, security/invariants, architecture/DRY) against the committed diff before push:

• Correctness: no P0s. One P1 — flagged that gmi-audio-clone may share the same defect but wasn't in the issue's confirmed reproduction; resolved by filing Confirm whether gmi-audio-clone (MiniMax Voice-Clone) also needs a prompt->text alias #257 rather than shipping an unverified alias.
• Security: no blocking issues. Two informational P2s (pre-existing shared alias-collision logging behavior in model_registry.py, and the same clone-family scoping note) — not regressions from this diff.
• Architecture/DRY: no blocking issues. Confirmed the ParamSurface.with_aliases() per-family composition here matches the existing video.py convention (e.g. _VIDEO_BASE/_PIXVERSE), adds zero per-request overhead, and is appropriately minimal.

Related

Refs #251 (intentionally not using a closing keyword — see the "Deliberately out of scope" note above; leaving the issue's closure to a maintainer's judgment call given the voice-clone open question)
Refs #257 (follow-up: confirm whether gmi-audio-clone needs the same aliasing)