Paul: "There's still a flash. Are you able to test this on a real browser?"
Yes — and doing so showed the previous branch fixed a real defect that was not this one.
What the flash actually is
A hidden tab closed its stream on the spot (noteHidden() { closeStream() }). Coming back reopened it and fetched the catch-up, and that round trip cannot beat the paint. The browser draws the document it already has — the version from before you left — and the corrected text lands a frame later. The paragraph visibly rewrites itself and reflows across lines.
On loopback the gap is 18ms. Over a real network it is a whole round trip, which is why Paul sees it plainly on the deployed host and why a local capture nearly missed it.
Nothing about making the reconnect faster removes this. The paint happens before any round trip can possibly return.
The measurement, both ways round
fixtures/flash-frames.mjs drives two real tabs: type in B, switch away, edit in A, switch back.
| | DOM has the edit while hidden | painted frames showing old text | |---|---|---| | before | false | 1 | | after | true | 0 |
The "1" is the flash, on screen, counted. The harness was proven able to see it — run against the old build first — before it was used to say it was gone.
The fix
A hidden tab shortens its deadline to 30s instead of closing. A switch away and back never disconnects, so there is nothing to catch up and the first painted frame is already current.
This serves §7.3 better than before, not worse: 30s is four times more aggressive than the 120s a merely idle visible tab is already allowed. A document left open on Friday drops its stream half a minute after you look away.
Two instruments that lied, both now recorded
Worth more than the fix, because both would have been believed again.
Counting DOM mutations measures a proxy. The previous branch drove 23 redundant attribute writes to 0 and reported the flash fixed. Those writes were real and worth removing — but a mutation count can go to zero while what a person sees is untouched, because a flash is a paint.
Chromium hands a screencast the last composited surface of a hidden tab. It arrives as an ordinary frame, indistinguishable from a paint. The first version of this harness read one as the flash — and it was delivered 787ms before the reveal, while the tab was provably still hidden. Adding a deliberate gap between starting the cast and revealing the tab is what exposed it.
The honest instrument is a requestAnimationFrame sampler: rAF does not run in a hidden tab, so its first entry is the first frame actually drawn. The screencast additionally only emits when the page happens to paint, and could not resolve a 20ms window at all — it showed an identical frame sequence for the broken and fixed builds.
Verification
Go 6/6 packages, JS 97/97 (two new tests, both confirmed to fail against the immediate close on the assertions that name the defect), wire-gate exit 0, browser drive 44/44.
amadan