Ads: generated spots render every tag in the station voice — the worker never writes a voice plan

[genwave-radio]

What airs

All 12 demo spots: ANNOUNCER, VOICE1 and VOICE2 in one voice. Kokoro log for spot #11 (2026-09-05 20:28:37–44Z): four POST /v1/audio/speech, all af_alloy. station.ad_spot.voice_plan is NULL on every generated row.

Why

• src/GenWave.Ads/AdRenderService.cs:123-127 — null/empty plan ⇒ every tag → stationIdentity.Current.Voice.
• src/GenWave.Tts/AdScriptWriteRequest.cs:44-51 carries no voice fields; AdScriptWriter.cs:397-398 says so: "for a generated spot every tag maps to the station voice anyway". Only the editor's voicePlan field or the API can set one.
• Per-line voice switching already works when a plan exists (CrosstalkAssembler.cs:208-216, 393-403), so this is a plan-building gap, not a render gap.

Ruling (Dean, 2026-09-06)

A separate ad cast pool, distinct from the DJ personas; the owner can set the ANNOUNCER voice (default Station:Voice).

Fix

• Settings: Ads:AnnouncerVoice (empty = station voice), Ads:CastVoices (list of Kokoro voice ids; ships with a small mixed default set distinct from the persona defaults). Live settings, help keys.
• Worker/refill: after the validator passes, build the plan — ANNOUNCER → announcer voice; VOICE1/VOICE2 → two distinct pool voices picked deterministically per spot (seeded by spot id + brand so a regeneration re-casts the same way), never equal to the announcer; persist as voice_plan (AdVoicePlanEntry(Tag, VoiceId, Pace) jsonb). Owner drafts with an explicit plan keep it.
• Editor/rows: show the cast as read-only chips so approval sees who says what; a cast picker is a later rider (Ads: the spot editor's Brief field is a dead end, and briefs want a "Suggest a brief" button #691 territory).

Acceptance

Kokoro's log for a rendered spot shows ≥ 2 distinct voices when the script has ≥ 2 tags; voice_plan non-null on worker-written rows; SPEC F161.2 rider.

Epic: #380