fix(tts): prevent runaway GPT-SoVITS filler vowels

[Lucas1479]

中文摘要

给中文审阅者的速览；下方英文正文保留完整实验数据。

• 问题：GPT-SoVITS 在静态 KV cache 和 CUDA Graph 路径下可能读取尚未写入或仅用于对齐的缓存位置，导致语义 token 坍缩为长串重复元音，最终产生失控的填充音。
• 修复：屏蔽无效 KV 位置；从真实 prompt 长度连续生成；bucket 溢出时保留完整条件上下文；在进入 SoVITS/BigVGAN 前拦截与 token ID 无关的坍缩候选，并只允许一次有界、可观测的重试；同时隔离 Graph capture 的随机数状态并使修复前首句音频缓存失效。
• 实验结果：固定 660 次、同 GPU/模型/文本/种子的对照中，最终 Graph+static、static、dynamic 三条路径的异常均从修复前的 24/220、24/220、14/220 降为 0/220；220/220 组语义流与保护决策保持一致。示例语音从 9.77 秒、最长 192 个重复 token，改善为 5.04 秒、最长重复 10 个 token，ASR 能保留完整续句。
• 性能与审阅重点：Graph/static 中位延迟仍为 108.4 ms，约比非 Graph 路径快 2.75 倍。建议重点审阅保护规则的误杀风险、单次重试的可观测性、随机数隔离以及缓存版本失效是否完整。
• 验证：50 项公开 TTS/backend 定向测试、语义保护 fail-closed 测试、第三方来源门禁、源码发布门禁、Python 编译与 diff 检查均通过。

关联问题：#21。

Summary

• mask unwritten and CUDA-Graph alignment-gap KV positions while preserving fixed graph shapes
• generate contiguously from the real prompt length and preserve full conditioning context on bucket overflow
• reject token-ID-independent collapsed semantic candidates before SoVITS/BigVGAN, with one bounded observable retry
• isolate graph capture from sampling RNG and invalidate pre-fix first-sentence audio-cache identities
• add paired-seed stability probes, real-audio rendering support, static/dynamic parity tests, and guard/cache regression tests

Evidence

A fixed 660-trial filler-heavy paired-seed cohort (same physical GPU, model hashes, texts, seeds, and 400-token cap) changed from:

Phase	Graph + static	Static	Dynamic
Before	24/220	24/220	14/220
KV fix only	14/220	14/220	14/220
Final guarded path	0/220	0/220	0/220

After the KV fix, Graph/static and static/dynamic semantic streams match exactly for 220/220 paired case/seed combinations. The final guard makes the same retry decision in all 220/220 pairs.

For うーん……正直、まだ完全には分からないわ。 at seed 2001:

• before: 243 semantic tokens, longest equal-token run 192, 9.77 s WAV, ASR only うーん。
• after: 125 tokens, longest run 10, 5.04 s WAV, ASR retains the complete continuation

Normal median T2S latency on the final full matrix is 108.4 ms Graph/static, 300.0 ms static, and 297.8 ms dynamic; Graph remains about 2.75x lower-latency than non-Graph paths.

Validation

• 50 passed targeted public TTS/backend regression suite
• direct project-environment semantic guard retry/fail-closed tests passed
• third-party provenance release gate passed
• deterministic source-release gate passed (0 errors, 0 warnings)
• Python compilation and git diff --check passed

Closes #21

[Lucas1479]

Maintainer self-review complete: no blocking findings. The decoder masks unwritten/alignment slots, writes from the real prompt length, and preserves full conditioning by switching to dynamic KV at bucket exhaustion. The pre-vocoder guard is token-ID independent, retries at most once, and fails closed; Graph capture no longer advances sampling RNG; the first-sentence cache revision prevents reuse of pre-fix audio. Validation on the rebased head in the formal .venv_cu124 environment (Python 3.12.10, Torch 2.5.1+cu124): 59 focused TTS/cache/Graph/guard tests passed, changed-surface Ruff/compile/diff checks passed, provenance passed with 0 errors, and deterministic source-release passed with 0 errors/0 warnings. A fresh physical-GPU paired probe (2 seeds x 4 quick cases x 3 modes = 24 trials, production one-retry guard) completed with 0 anomalies, 0 errors, 100% EOS, and 100% Graph/static and static token parity against dynamic. Required build and cpu-model-less checks are green on head 7bb99b3. Admin merge bypass is used only for the unavailable external approval requirement.