[Lucas1479]

What and why

Fixes #235.

The Windows ROCm profile uses TorchAudio 2.9, whose load() requires TorchCodec. A swallowed reference-cache failure let v3 synthesis continue without its semantic prompt; the semantic guard then misread the decoder's reference-free index as an empty result, and streaming reported synthetic silence as success.

• Decode local v3 references with the existing SoundFile dependency and reuse only complete reference caches.
• Slice generated tokens after the actual prompt prefix, preserving valid v1/v2 reference-free synthesis and the existing bounded semantic guard.
• Propagate live preparation and streaming errors instead of yielding silent success.
• Advance the first-sentence synthesis revision to v5 so existing silent v4 entries are ignored, including legacy asset-layout entries with or without metadata. Existing files remain on disk; audio regenerates on demand.
• Keep optional emotion-reference warmup best-effort for ordinary exceptions, with an invalid status and warning; default synthesis remains available.
• Integrate main at 559e9da0c63a899b40eef1727634a1177332c850. The ROCm workflow retains both fix(tts): load v3 references without TorchCodec and preserve failures #238 reference/semantic suites and fix(tts): serialize shared playback stop and initialization #242 playback recovery coverage in its triggers and test command.

Change class

• Routine fix

Owning layers: local GPT-SoVITS reference preparation and semantic-result consumption, the existing derived-audio cache revision, and the optional reference-pack startup boundary.

Compatibility: no dependency, public API, configuration, or model-format change. First-sentence derived audio is regenerated once under the new synthesis revision. Live failures retain their original exception.

Evidence

Current head: 87e6c00faf23e4217995d86211d83345f385ff94.

Windows / Python 3.12.10; torch and torchaudio 2.6.0+cu124 using CPU tensors, SoundFile 0.13.1, librosa 0.11.0:

AMADEUS_E2E_NO_TTS=1 TTS_DEVICE=cpu WORK_WORKTREE_ISOLATION=0 HF_HUB_OFFLINE=1
python -B -X utf8 -m pytest -q -p no:cacheprovider
  tests/test_local_tts_semantic_guard.py tests/test_local_tts_reference_cache.py
  tests/test_t2s_semantic_history.py tests/test_tts_semantic_stability.py
  tests/test_tts_emotion_references.py tests/test_local_tts_punctuation.py
  tests/test_tts_contract.py tests/test_gpt_sovits_sidecar_backend.py
  tests/test_first_sentence_audio_cache.py tests/test_playback_device_recovery.py

• 174 passed, 1 skipped; punctuation tests are gated off in model-less mode.
• The seven new/extended cache and optional-warmup cases failed before the follow-up fixes and pass afterward. Coverage includes both cache identities, entries with/without metadata, regeneration and subsequent hits, and IndexError/ImportError during optional warmup followed by successful default synthesis.
• Original reference/semantic regression evidence at unchanged 7b97e21c: 25 failed, 7 passed; the fixed source passes all 32 cases.
• Reference tests use synthetic audio with real CPU decoding, resampling and mel extraction. Neural network outputs are stubbed; no checkpoints or character assets are required.
• Ruff on the four follow-up Python files, git diff --check, and merged workflow YAML validation passed.

All 16 PR checks passed on this head, including Electron build, Windows packaging/startup smoke, the full model-less aggregate, Linux/macOS voice, cu124, Windows/Linux cu128, Linux T4 CPU, and macOS MPS.

• Full Windows model-less CI: all three isolated shards passed; 5,974 passed, 36 skipped, summed from the runner pass/skip summary columns.
• Windows ROCm candidate CI: 175 passed, 1 skipped on the official torch/torchaudio 2.9.1+rocm7.2.1 environment. Hosted runners use CPU tensors and mocked playback devices.
• Local model installation candidates and the cu124 single-environment ladder passed.

Not yet verified: real Windows AMD ROCm 7.2.1 local v3 inference, GPU/driver behavior, or speech quality. #241's successful AMD Fish Audio playback report validates a different backend path; it does not validate this local v3 inference change. Hardware feedback remains welcome and is not required for merging this bounded code fix after current-head CI passes.

Final check

• One coherent defect and its cache/startup consequences; no unrelated feature work.
• Existing contracts reused; incomplete-cache fallbacks removed.
• No secrets, local state, model weights, or voice assets included.
• Vendored model code and third-party provenance unchanged.