Record six traps from a day of contended CI and self-inflicted churn

[samoylenkodmitry]

Follow-up to #510, opened after it landed so this one gets a clean run. Six entries, one push, deliberately frozen — #510 took nine pushes and sixteen cancelled runs because good material kept arriving, and that is itself one of the entries below.

• Within-run control before believing a perf regression — an untouched stage moving the same direction means you measured the governor, not the patch (from Scope a lazy-list scroll to its subtree instead of rebuilding the whole scene #504's near-miss).
• Knowledge PRs have a convergence budget — and the check rollup hides the runs you already killed; gh run list shows them.
• A cache key can be right about what it identifies and wrong about granularity — the backdrop key takes screen-space bounds in a local_bounds slot, and the obvious translation-invariant fix would serve stale backdrops that look correct on a slow scroll.
• Closing a PR does not cancel its CI — a closed PR started a robot job three hours later on one of only two machines that can run it.
• gh pr merge can exit non-zero on a merge that succeeded — only the ancestry check settles what landed.
• Test the accept path — a gate tested against seven bad inputs held on all of them and could never fire, because its one untested input was the green case.

Two further items are deliberately not here: the robot wait-string disambiguation is gated on #508 and the a11y_sync toggle citation on #497. Neither has landed, and an entry quoting text that is not in the tree a reader is holding is worse than no entry.

Attribution: the within-run control and cache-key items are others' observations, verified against the tree before writing (raster_cache.rs:114 takes local_bounds, render_paths.rs:3032 passes visible_rect). The last three are my own failures from today.

🤖 Generated with Claude Code