Two gates per scene put about thirty confirmations into a fourteen-scene run

Scene count is the only input to the gate count. The post labels the pattern 2N+2, though the stops it lists (script, board, two per scene after the first, final) add up to 2N+1, which is 29 on a fourteen-scene run ^[23]. At fifteen scenes the label gives 32 ^[22]. Either count puts the editor near thirty clicks, and each extra scene costs two more.

That linear scaling is what made oversight bimodal. A frame gate on scene nine costs the editor exactly what the script gate costs, one more stop, and catches much less. Skipping all of them was the only other setting the tool offered ^[6], and how often editors chose runAll is not in the post ^[27].

The fix keeps four stops, and keeps the ones that gate money ^[7]. For four to be the right number in another pipeline, per-scene failures have to be correlated or cheap to redo. A bad board that reliably produces bad frames is covered by one gate at the board. Independent clip failures escape that: at four to six minutes a clip ^[21], fourteen scenes is 56 to 84 minutes of rendering ^[26], and batching review to the end means paying for all fourteen to find the two that are wrong.

The create form's flags are the best piece of engineering in the set. They resolve through a dependency graph, and a flag whose parent requirement is unmet resolves to false no matter what the operator set ^[12]. Six of them therefore read "On" and changed nothing about a paid generation, about one control in nine ^[25]. The client now mirrors the same transitive resolution, where before it decided the question on its own ^[14]. Two copies of one rule will drift. A second, independent opinion about what a flag means would drift faster and be harder to notice.

Grouping landed in PR #879: 32 of 32 settings carried a group label and none of the 56 flags did, with the component's own code comment conceding the gap before anyone fixed it ^[15][16].

The dashboard fix reused what was already there. groupRunRows already computed which runs were waiting on a person, oldest-wait-first, and was already called elsewhere in the app. The landing page had never called it. It sorted newest-first and showed five, so a run parked at a gate could sit below three finished runs and fall off the bottom of the strip ^[17][18]. PR #964 replaced the sort with two headed blocks, "Waiting for you (N)" and "Everything else" ^[19].

The click-through audit in PR #865 found three dead-ended paths ^[8]. The button offered after a run fails built its URL against a route deleted three weeks earlier, so the recovery path from a failure was itself a 404 ^[9]. The diagnostics page had one inbound link, gated on a benchmark-model field the create form had stopped writing. No run the app could currently start would open it ^[10]. The library section (Products, B-roll, Avatars, Profiles) rendered only while its own section was already active in the URL. That left four destinations invisible from anywhere else ^[11]. And scenes 0 through 2, rendered and billed, hid behind a loading panel whenever the focused scene had no finished take ^[20].