Reshoot the demo video so it shows what actually shipped - #53
Merged
Conversation
added 3 commits
August 3, 2026 19:10
The committed cut was 2:56 and predated four things a judge should see: the guided tour, the case id in the URL, the measured render estimate, and the plain-language provider failure. Reshot against the live Cloud Run app at commit ba00f8e (checked via /health build.commit before shooting). Three new cards, captured as a guest at a 1920x1080 viewport with deviceScaleFactor 2 so the tight crops stay sharp: 14-guided-tour the hero with the opt-in tour open at step 1 of 6 15-measured-wait the review step's estimate, in its own words: 'about 2 minutes ... from the last 1 case rendered here' 16-case-link case s05_t_intersection-3244ce14 REOPENED from its own #case/ link, showing the link card and the full address The third one is the point of the beat: that card exists because the resume path works, not as a mock of it. No render was commissioned for this pass. The tour needs no case, the estimate needs only an extraction, and the sealed case was fetched back rather than remade, so the reshoot cost one vision call and no generative credit, which matters on a day the provider ran dry once already. Two beats added (03a-tour, 09b-link) and 05-render moved off the async-progress screen onto the measured wait. The cap is 180s and the old cut already spent 176.2s, so every remaining line was tightened to pay for them: the problem beat lost a third of its words, the authorship beat lost the half that repeated the storage beat, and the accuracy beat lost its preamble. Final cut is 172.600s over 15 beats, 7.4s under the cap. The provider failure is narrated, not shown. A real one cannot be induced against the live app on demand, and a mocked screenshot would break the rule that every card in this video is the running product. The favicon is not in it at all: a page capture cannot show browser furniture, and drawing a fake tab strip would be the same lie. Both decisions are written down in the beat script's description rather than left for someone to notice. README and demo/SUBMISSION.md now say 2:53 / 172.6s and fifteen beats, with the per-criterion budget recomputed from the built windows.
Both were run once from an owner's shell and were invisible to the next reader, which is exactly how the Hosting site went missing in the first place: adding Firebase to what was still a bare GCP project, and granting the existing Cloud Run deploy identity roles/firebasehosting.admin and nothing else. They now sit in the workflow header next to the trust path they belong to. Creating the Hosting site is deliberately NOT on that list. The deploy does it itself, so it cannot be forgotten again.
This job hung inside apt-get twice this afternoon, thirteen minutes each time, with no output and no failure. A hung gate is worse than a failing one: the Cloud Run and Firebase Hosting deploys both fire on CI's completion, so a stall here silently stops both while still looking like work in progress. Three changes, all small: a ten-minute job timeout (many times what the job needs, so it only fires on a stall), a skip when the runner image already carries ffprobe, and apt retries so a flaky mirror is retried rather than sat on.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The committed cut was 2:56 and predates four things a judge should see. Reshot against the live Cloud Run app at commit
ba00f8e, confirmed through/healthbuild.commitbefore shooting.What is new on screen
03a-tour(new)14-guided-tour05-render(recut)15-measured-wait09b-link(new)16-case-links05_t_intersection-3244ce14reopened from its own#case/link, link card and full address on screenThe third card is the point of that beat: it exists because the resume path works, not as a mock of it.
What the reshoot cost. Nothing generative. The tour needs no case, the estimate needs only an extraction, and the sealed case was fetched back rather than remade, so this was one vision call and no render credit, on a day the provider had already run dry once.
What paid for it. The cap is 180s and the old cut already spent 176.2s, so every remaining line was tightened: the problem beat lost a third of its words, the authorship beat lost the half that repeated the storage beat, and the accuracy beat lost its preamble. Final cut 172.600s over 15 beats, 7.4s under the cap. No beat is padded beyond the fixed 0.35s tail, so nothing freezes:
Two things deliberately not shown. The plain-language provider failure is narrated, not shown: a real one cannot be induced against the live app on demand, and a mocked screenshot would break the rule that every card here is the running product. The favicon is not in the video at all, because a page capture cannot show browser furniture and drawing a fake tab strip would be the same lie. Both are written into the beat script's description rather than left for someone to notice.
README and
demo/SUBMISSION.mdnow read 2:53 / 172.6s and fifteen beats, with the per-criterion budget recomputed from the built windows.