Ads first contact: the prompt's literal "TAG" is copied by the 3B model, and the re-ask shares the timeout budget — zero usable spots on the demo

[genwave-radio]

First contact, 2026-09-05: v5.6.1 on the demo, Brought To You By installed (30 briefs, all enabled), AutoApprove off. Two worker ticks, zero usable spots. Both failures are the app's, not the model's.

🐛 1. The prompt's abstract "TAG: <line>" is copied literally

AdScriptPromptBuilder.BuildSystemPrompt says: in this EXACT wire format … "TAG: <line>". llama3.2:3b (the reference station's model, HARDWARE.md) copies it — the first draft of BOTH ticks opened with a line whose tag was "TAG (leading quote included), refused by AdScriptParser's tag rule:

Ad script failed (brand: Pruett's Left-Handed Hardware): voice tag ""TAG" is not uppercase-alphanumeric, starting with a letter
Ad script failed (brand: Lumbago Larry's Discount Mattress Barn): voice tag ""TAG" is not uppercase-alphanumeric, starting with a letter

The crosstalk prompt never has this problem because it spells its example with the REAL tags — "HOST: <line>" or "NEIGHBOR: <line>" — so a verbatim copy is a valid line. The ad prompt must do the same: name real tags (ANNOUNCER, CUSTOMER, VOICE1), state the tag grammar in words (one word, capitals and digits, no quotes, no parentheses), and drop the abstract placeholder. Bench on the demo's exact model to follow in the fix PR.

The second tick's re-ask then produced LARRY (YELLING) — a stage direction inside the tag — and landed as the visible failed row, exactly as designed (STORY-390 AC3). The tag grammar sentence covers that one too.

🐛 2. The re-ask shares one Llm:TimeoutSeconds budget with the first attempt

AdScriptWriter.WriteAsync: ONE timeoutCts spans both attempts ("mirrors LlmCopyWriter's re-ask callers"). That mirror is the wrong shape here: the on-air writers share a budget because the break is imminent; the ad writer runs entirely off the air clock (SPEC F161.1, the class's own remark for why it carries no single-flight gate). On the demo's CPU-bound 3B model a first completion took 49 s of the 90 s budget, so the re-ask timed out and the tick produced nothing:

Ad script failed (brand: Pruett's Left-Handed Hardware): Llm:TimeoutSeconds exceeded
Ad script generation skipped for brand Pruett's Left-Handed Hardware: Llm:TimeoutSeconds exceeded

Fix: each attempt gets its own Llm:TimeoutSeconds budget in the ad writer (the worker's 10-minute tick is the only clock that matters). One spec: a first attempt that spends most of the budget on a refused draft, then a re-ask that itself fits the budget, succeeds.

📉 Why it matters now

At the current shape the demo's stock pass yields ~0 spots per tick, so T409's ear gate never starts. Related: #693 (the structure hint never reaches the prompt — the same builder). Epic: #380.

[genwave-radio]

📊 Bench on the demo's exact model — llama3.2:3b, ollama default sampling, same system/user shape as AdScriptWriter

Six pack briefs, N runs each, scored with the validator's own rules (tag grammar, ANNOUNCER required, 1–3 voices, per-line cap, 30 s ± 40 % at 15 chars/s, 555 phones). Two independent samples.

Prompt wording alone (sample 1, N=4 per brief, 24 runs per variant)

variant	pass	dominant failures
current ("TAG: <line>" placeholder)	7/24 (29 %)	stray untagged lines 8, no ANNOUNCER 5, LARRY (YELLING) 2, literal "TAG 1
prose "one word in capitals, like ANNOUNCER…"	2/24 (8 %)	Announcer/Voice1 mixed case, VOICE 1, the brand as a tag
prose + two example lines	5/24 (21 %)	no ANNOUNCER 9, more than 3 voices 4

Even a "pass" under the current wording used the literal TAG as a voice (TAG: GET READY FOR THE BIGGEST SALE…), which the tag grammar accepts.

Crosstalk-shaped prompt + a normalising pass before the validator (sample 2)

variant	raw pass	after normalising	residual after normalising
current placeholder (N=2, 12 runs)	0/12 (0 %)	11/12 (92 %)	too long 1
"ANNOUNCER: <line>" … "VOICE1: <line>" or "VOICE2: <line>" (N=4, 24 runs)	19/24 (79 %)	21/24 (88 %)	too long 1, too short 1, >3 voices 1

The normalising pass folds exactly the shape quirks seen: wrapping quotes, mixed-case/spaced/apostrophed tags, a parenthetical inside the tag, a beat label used as the speaker, a tag on one line with its words on the next, title/bracketed/header lines, and a cast with no ANNOUNCER (the lead voice becomes it — every tag maps to the station voice on a generated spot anyway). What survives it is real content the re-ask exists for: length and cast size.

Fix PR carries both halves plus per-attempt timeout budgets.