Rebuild the film, and take the test count out of the narration - #59

The two films on disk were built on 2026-08-13 — ten releases back. Rebuilt against today's tool, and the one line that could not rebuild itself was re-recorded.

The screen fixed itself; the voice could not

scene_kit.test_count() reads the suite size off the README instead of having it typed into a slide. Its own docstring is the argument:

Typing it into a slide would put a number on screen that nothing forces to stay true — the exact failure this material is about.

So 08-roslyn moved 295 → 556 by itself, in both languages. That was the entire diff to the scenes — everything else re-rendered byte-identical, which is what the deterministic capture is there to buy.

But a recorded take has nothing keeping it honest, and the narration said "doscientos setenta" / "two hundred and seventy". The rebuilt film was about to contradict itself out loud.

The number is gone from the line rather than corrected to 556. Correcting it buys a stronger sentence that is wrong again next release and needs a re-record every time the suite grows; dropping it costs one re-record, ever. The count stays on screen, read from the one place with a test enforcing it.

es  "...no hace falta licencia, y sus tests corren gratis en un runner público."
en  "...it needs no licence, and its tests run free on a public runner."

Re-recorded with the voices the guion already names — Marcela for Spanish, Rachel for English, who is not in the account and was already pinned by id in PREMADE_VOICE_IDS for exactly this reason. The key was passed through the environment for the run; nothing was written into this repository, and the diff carries no credential.

The shorter line makes the scene shorter: 17.8s → 17.0s (es) and 16.4s (en). --retime read that off the new takes, every other scene came back at the length it already had, and the six chapter stamps after 08-roslyn in PUBLICACION.md moved with it.

The evidence stills had gone stale, and checking that is the point

ensure_stills captures the report sections the evidence scenes pan over, and only ever captures when the file is missing. They sat sixteen days and ten releases behind the tool output they are supposed to be, and nothing in the pipeline would have said so.

Deleting them and recapturing produced byte-identical files — the fixture still extracts the same, so the film was never showing anything untrue. That was luck rather than design.

--restills makes it a command instead of an assumption:

python docs/Blog/video/build-scenes.py --restills --video 07-pantallas 10-mapa

A byte-identical recapture is itself the useful answer.

One thing left for you, deliberately

The blog article says "270 tests" in four files — xaflogic-hidden-behaviour-story.{es,en}.{md,html}.

I did not touch it, because it is a different kind of claim. A film is regenerated from live sources, so a stale number in it is a defect. An article is a dated publication, and what it said when it was published is not false. Updating it to 556, rephrasing it the way the narration just went, or leaving it as a snapshot are all defensible — which is why it is a decision rather than a fix.

Also checked and still true: "catorce clases" (14 entity classes in PharmacyDemo.Module), and no other hard figures are spoken in either language.

render/ stays untracked — 36 MB of derived output. The finished films are on the machine that built them.

🤖 Generated with Claude Code

https://claude.ai/code/session_01W45tzJFX3NoSrk7svtQeKT