fix(tts): ban commas in LLM patter — the prosody stumble at its source (gh-#303)

[genwave-radio]

Closes #303. Relieves #292.

🎧 The fault

Commas are the single biggest source of unnatural pauses on air. A small model writes grammatically correct clause-heavy copy, and both engines honor every comma with a stumble the ear reads as hesitation. #292's comma-before-vocative (hats, folks) is the same fault at its most audible.

Overnight sample from the demo box — every one of these is comma-led:

"That was an infectious start to the day with 'Love Will Come Back Again' by Chaz Robinson, getting e…"
"Getting our day underway with a taste of Havana, we're just seconds away from…"
"Here's your daily dose of optimism, and we're kicking off with a track that…"

🛠️ The fix

Four sentences added to the shared scaffold body in LlmPromptBuilder.BuildSystemPrompt, so persona-voiced and persona-less copy are held to it equally.

Two escape hatches, deliberately different:

the model wants a…	it must write…	what the ear gets
real clause break	a new sentence	a true 0.6s silence — #116 inserts [pause:0.6s] at sentence ends on the Kokoro path
run-together phrase (hats folks)	the words, no comma	speech that runs through — #292's ask

Collapsing both into "always split" would trade a 0.2s stumble for a 0.6s gap and read worse for exactly the vocative case. That's why the rule is two rules.

The rule states itself without commas, because prompt text is style the model imitates. A spec locks that in — it's the kind of intent a well-meaning future edit quietly undoes.

🚫 What this deliberately does NOT take

The issue's second half — "we should be using [pause:0.6s] explicitly instead of relying on a side effect":

1. CleanCopy already strips every bracketed run (\[[^\]]*\]) as a stage direction, so a model-emitted tag never reaches an engine today. Taking this would need a deliberate carve-out in that filter.
2. piper-tts speaks any tag aloud (TtsOptions.cs:29) — and the Piper hop is exactly what runs when Kokoro dies (tts: kokoro dies silently mid-render (likely OOM) — Piper fallback voice airs on demo #276). A carve-out would put "bracket pause zero point six s" on air on the fallback path.

Sentence boundaries + #116 deliver the same explicit silence, with the placement decided by us instead of by a 3B model. Happy to revisit if you want author-controlled pauses in human-written copy (persona cards, templates) — that's a different, safer surface.

✅ Verification

• dotnet build clean, 0 warnings
• dotnet test GenWave.sln — 2,547 passed, 0 failed (Tts specs 329 → 335)
• No existing spec asserts the scaffold text verbatim; llm: patter confabulates artist facts and parrots quirk example values — tighten the system prompt's embellishment license #188/llm: patter must use they/their for artists — never infer gender from a name #151/llm: 'personality-neutral' boilerplate contradicts persona Style in the same system prompt; consider exposing the prompt in the Admin UI #152 substring guards all still pass

Prosody itself is an ear check — worth listening to a few breaks after this ships.