fix(desktop): make render-churn measure streaming, not boot churn

[OutThisLife]

Trust-the-instrument work before the next perf fix. Brooklyn pushed on whether the harness was actually valid given the gateway errors scrolling past every spawn — the errors turned out to be harmless, but checking found the harness itself was not trustworthy.

The measurement was mostly noise

Three back-to-back runs, identical code:

run	total_renders	wasted_renders	commits	sidebar
1	19,857	4,192	1,033	6
2	43,294	16,002	646	12
3	20,576	4,639	917	6

2.2x spread on renders, 3.8x on wasted. Wide enough that a single-run before/after delta is as likely to be measuring machine load as a code change.

Cause: the scenario slept a fixed 1s after mounting tabs and then recorded. Session hydration isn't reliably finished by then, so a variable amount of boot work landed inside the window — which is why the bad run's top offenders were Primitive.div and TooltipProvider rather than anything on the streaming path.

Replaced with a quiesce gate: poll the commit counter until it holds still, then record. It reports quiet:N (comparable) or timeout:... (contaminated) so a bad run is visible in the output instead of quietly skewing a number.

The counter was mis-attributing

wasted counted any render with no changed props and no changed hook state — which lumped in context-driven renders. That points at memo() as the fix, and memo() cannot block a context update at all. Adds contextChanged off the fiber's context dependency list and excludes it from wasted.

This immediately disproved my own working theory: Block/Ct come back props=0, state=0, ctx=0, so the streamdown context I suspected is not involved.

The finding that matters more than the fix

With five busy tiles and no driver running, the renderer still commits ~18x/sec. Block/Ct render 800 times while completely idle — more than during the actual streaming window. So the transcript cost I was about to "fix" is substantially an idle-churn problem, not a streaming one.

The report now names cascade roots (own state changed, props did not) instead of leaving them to be guessed:

GlyphSpinner               r=193  state=193
Streamdown                 r=105  state=105   <- parent of Block/Ct
ResponseLoadingIndicator   r= 74  state= 74

Streamdown re-rendering itself 105 times while idle is what drives Block/Ct. That's the next PR — this one just makes it measurable.

No behavior change; instrumentation and harness only.

[github-actions]

૮ >ﻌ< ა ci review

_{ran on f7aee9d}

all good!