fix(ads): the writer's prompt names real voices, normalises the reply, and budgets each attempt on its own (#696)

[genwave-radio]

Closes #696. First-contact fix for the ad writer, three halves, each pinned; bench evidence on the issue.

🔧 What changed

1. The prompt spells its example with the real tags (AdScriptPromptBuilder): "ANNOUNCER: <line>" for the announcer, "VOICE1: <line>" or "VOICE2: <line>" for up to two other voices, plus one sentence of tag grammar in words ("the word before the colon is always the VOICE speaking … no quotes, no parentheses, no stage directions; never a beat name … never the brand name"). The abstract "TAG: <line>" placeholder is gone — the reference model copied it verbatim, quotes included. This is the CrosstalkPromptBuilder shape, which is why that writer never had the problem.
2. A normalising pass before the validator (AdScriptWriter.ApplyLineAwareHygiene, LLM path only — owner text stays verbatim per F160.4): wrapping quotes/bullets/emphasis stripped from the whole line before the tag is split; the tag folded to [A-Za-z0-9] upper-case with parentheticals dropped (Announcer, VOICE 1, PRUETT'S, LARRY (YELLING)); a beat label used as the speaker (HOOK:, Tagline:, the literal TAG:) is the announcer; an untagged line joins the previous voice (filling a bare TAG: when the model put the words on the next line); title, [bracketed]/(parenthesised) and # lines are dropped; and when no line is tagged ANNOUNCER, the most frequent voice is. Never the chat-preamble heuristic on a tag — the T400 review's hazard stays closed. A still-empty tag stays visible for the validator, as before.
3. Each attempt gets its own Llm:TimeoutSeconds budget (NewAttemptBudget): the on-air writers share one across a re-ask because the break is imminent; this writer runs off the air clock, and on the demo's CPU-bound model one completion took 49 s of the shared 90 s, so the re-ask timed out by construction.

📊 Why this shape (llama3.2:3b, the demo's model — full tables on #696)

	raw pass	after normalising
current prompt	0–29 % across two samples	92 %
this prompt	79 %	88 %

Residuals after both are real content for the re-ask: too long, too short, more than three voices.

🧭 Two design calls inside, flagged for the ruling

• The lead voice becomes ANNOUNCER when the model cast nobody as one. For a generated spot every tag already maps to the station voice (AdRenderService, null voice plan), so nothing audible changes; it turns a whole re-ask into a rename. Drop the rule if you'd rather the re-ask carry it — it is one block.
• A leading untagged line is dropped (a title, a chat preamble) rather than left for the format rule. Same trade: 3B models title their work; the re-ask was paying a full completion to delete a title.

✅ Verified

• dotnet build GenWave.sln: 0 warnings, 0 errors.
• Tts 857 passed (11 new: prompt shape, 9 normalisation rules, per-attempt budget — the budget spec proven discriminating by mutation: shared budget → exactly one red), Ads 98, Architecture 135. Host 2565 passed with one red — the real-Postgres migration probe, which ran while I was removing the bench's ollama container; it passed on an isolated rerun (1 m 23 s).
• Mock server gains a ServeDelayMs knob (a slow success, distinct from the hang mode) for the budget spec.

📻 What the demo did meanwhile

Four ticks, four distinct shape quirks, zero usable spots — every one is a row in the normaliser's table:

tick	first draft	re-ask
1	tag "TAG (quotes included)	timed out (shared budget)
2	tag "TAG	LARRY (YELLING) → failed row
3	VOICE OVER 1	timed out (shared budget)
4	ANNOUNCER alone on a line, words on the next	an untagged line → failed row

Patch-release candidate: the demo stays at zero usable spots per tick until this lands.