Yueqian Lin
1ad90653c5
dots.tts landed in #4765 with no GPU coverage: nothing under tests/e2e exercises it, so the DiT sampler, streaming vocoder and AR loopback have no CI signal at all. Add an offline e2e module marked slow + tts + L4, which the existing weekly 'TTS · L4' step collects via -m 'slow and L4 and tts'. No pipeline YAML change is needed; the weekly TTS steps are marker driven. Three cases, the latter two guarding specific review findings from #4765: 1. zero-shot synthesis produces finite, non-silent audio 2. mixed prefill+decode batch keeps per-request logits rows aligned 3. the same text twice yields the same audio (seeded FM noise) Verified selector behaviour: 3 collected under the weekly selector, 0 collected under the nightly one. Signed-off-by: Yueqian Lin <linyueqian@outlook.com>