test(browser): gate warm-switch timing on the run median

[loganj]

🤖

Summary

• A browser speed check failed on main (run) because one of four channel switches took 206 ms against a 200 ms limit. The other three took 131–183 ms, and the commit did not touch channel switching.
• The check now applies the 200 ms limit to the median of the four switches, not to each switch. One slow switch no longer fails CI. A real slowdown in switching still does, because it slows every switch.

Why the median

• The hosted CI machines differ in speed by about 2x. On 21 main runs (Sep 30 – Oct 1), the per-run median ranged 90–176 ms and the cold-open time ranged 256–517 ms. The failed run was the slowest machine in that set: its cold open (517 ms) and all four switches were high together.
• With the old rule, 1 of those 21 runs failed. With the median rule, all 21 pass. This PR's own CI run gives another example: its switches took 334, 195, 167 and 171 ms, so the old rule would have failed, while the median (183 ms) passes. A ceiling set to 10 ms still fails, which shows the gate is still active.

Is it a regression?

• Not from the change under test. On one machine, current main and a commit before Unify workspace panels and add persistent channel tabs #413 had the same median (47.5 vs 47.8 ms, 72 samples each).
• Over a longer period, switching did get slower. Locally, the median went from 35 ms (Sep 23, when this limit was set) to 48 ms now. The cost came in small steps. For example, Keep pages and channels in a persistent sidebar #234 (persistent sidebar) added about 3 ms. No single commit caused it. This PR does not fix that drift. The per-switch 100 ms target notes still record it for performance review.

Details

• Every sample stays in evidence.json and in the target-miss notes. The run median is also recorded there.
• Remaining risk: the machine-speed spread is still about 2x. The worst median seen so far (183 ms, on this PR) leaves only a small margin, and a slower machine could still fail the check. Only large slowdowns will fail CI. The 100 ms target notes are the signal for smaller ones.

[loganj]

Closing. The switch got slower over time, so the fix belongs in the app, not in the test.