Show when an ad was re-rendered automatically

[genwave-radio]

❓ What

The #854 stale pass re-renders every ready spot with render_version < Current. That includes spots promoted from an auditioned preview, so the take the operator approved by ear is replaced by a new render they haven't heard. It uses the same script, cast and bed, but it's a different performance.

🤔 Options

• Keep it: the point of a version bump is that old pipelines sounded worse, so every spot should get the fix.
• Skip promoted previews: stamp them at the current version on promote, or exclude them in FindStaleReadyAsync, and let the operator re-render by hand.
• Ask the operator: flag them in the console as "a newer render is available".

✅ Done when

• There's a product decision (Dean), and the behaviour matches it, with a spec.

Follow-up to #854 (PR #861).

[genwave-radio]

🔬 Kokoro repeatability, measured 2026-09-26 (dev box, kokoro-fastapi-cpu:v0.6.0)

I rendered the three lines of dev spot 10 twice each. I used its real cast (af_heart, am_michael, af_nova) at pace 1, calling /v1/audio/speech directly, then aligned the two takes and compared them.

Voice	Δ duration	Lag	Correlation	SNR	Level Δ
af_heart	0.1 ms	+0.12 ms	0.994	19.0 dB	0.01 dB
am_michael	0.0 ms	0.00 ms	0.992	17.7 dB	0.01 dB
af_nova	0.8 ms	−0.12 ms	0.994	19.3 dB	0.01 dB

Takes are not byte-identical (three renders per line gave three different hashes). But the timing, the pacing and the level match to within a millisecond and a hundredth of a dB. The leftover difference is in fine waveform detail. My inference is that it comes from the decoder's noise component; I haven't confirmed that. The delivery isn't different.

✅ Recommendation: keep re-rendering promoted previews (option 1)

• What the operator approved by ear was the script, the cast, the voices and the bed. A re-render keeps all of those, and the delivery measures the same.
• Why the bump happens: it exists to fix pipeline defects (Ads: the last ~3 s of every spot is ducked under the lead-in — ads are annotated gw_tts=false, so the engine runs the music→voice transition on them #699 tail, ads: the background music stutters where a short clip repeats (aloop has no crossfade) #855 bed click, tts/ads: '(555) 0142' is read as cardinals and slips past the 555 rule (PhoneShape misses paren + 4) #856 digits). Skipping approved previews would keep known-bad audio on air, often on the spots the operator cared about most.
• Asking the operator on every bump would be a per-spot chore that mostly never gets done.
• Add one thing: show "Re-rendered automatically on vX" on the ad page, plus a booth-log line on swap, so a changed take always has a visible reason. The old file stays on disk (see Clean up ad recordings left behind by re-renders #864 for its eventual clean-up).

Awaiting Dean's decision.

[genwave-radio]

✅ Decision (Dean, 2026-09-26): option 1

Promoted previews keep being re-rendered on a version bump, which is the behaviour already shipped in v5.13.1. What's left for this issue:

• The ad page shows "Re-rendered automatically on vX" for a spot whose take was swapped by the stale pass.
• The booth log gets a line when a swap happens.

✅ Done when

• The swap records the version and time it happened. Both surfaces render it in plain language, with no jargon.
• There's a spec for each surface.