chore(test): De-flake slow and racy tests - #24452

Three unrelated flakes, all of which pass locally and fail only
under CI load.

**Tests that time out loading the SDK**

Several unit tests time out in CI on the first test in a file.
The first dynamic import of a large SDK graph pays for
transforming it, and that cost is charged to whichever test
triggers it. Locally it is a few hundred milliseconds. On a
loaded runner it crosses the 5000ms test timeout.

Swept every unit test in the repo and fixed the files that show
the pattern. Fixed with a bare side-effect `import` at the top of
the file. Collection is not bounded by the test timeout, so
pulling the load into it takes the cost off the test without
changing what the test exercises.

Exception: `nuxt vite/orchestrion` has to register its
`vi.doMock` calls before the module under test loads, and a
static import is hoisted above them, so it warms the graph from
`beforeAll` instead.

| file                                   | before | after |
| -------------------------------------- | -----: | ----: |
| browser-utils inp                      |  594ms |   9ms |
| browser-utils metricObserverOrdering   |  409ms |  22ms |
| nextjs appRouterRoutingInstrumentation |  750ms | 233ms |
| nuxt vite/orchestrion                  |  536ms |   1ms |
| sveltekit index.workerd                |  625ms |   8ms |
| tanstackstart-react client/tunnelRoute |  433ms |   1ms |

Which file pays is not fixed: the two browser-utils files traded
places between runs, because the transform lands on whichever
test reaches the graph first.

Three tests in these packages still run over 400ms, and all three
are slow for a reason:

- The two nextjs `appRouter` tests busy-wait 1100ms on purpose.
- `conflictingDebugOptions` re-evaluates three SDK entry points
  for each `debug-build` mock, which no warmup can avoid.

**INP span never sent**

`hidePage` waits for one `requestIdleCallback` before hiding,
assuming web-vitals has already queued its own idle callback and
will therefore run first. But web-vitals only queues its callback
once the click's Event Timing entry reaches its observer, which
is after `hidePage` queues its own. So `hidePage` wins the race
and hides the page too early.

That loses the interaction. The forced report web-vitals does on
`visibilitychange` runs while the metric is still -1, and the
pending callback that sets the real value afterwards reports
unforced. No INP is emitted at all and the test waits out its 30s
timeout.

Fixed by waiting for the entry and queuing the idle callback from
a task after it, which puts web-vitals' callback first. Callers
that never interacted skip the wait, so the 13 that only need the
page hidden pay nothing.

The react-router-6 e2e suite now passes 10/10 and drops from 1.4m
to ~53s.

**Cron job that never runs**

Two bugs in `suites/cron/cron`, both reproducible under CPU load.

`cron`'s `start()` sizes the delay to the next tick from two
clock reads, clamps a negative result to -1, and then treats that
-1 as "stop". A pause between the two reads that straddles a
second boundary is enough. The job is left scheduled for nothing,
and the scenario exits on its own 15s timer having sent nothing.

That is the reported failure, which reads as `0/5 expected
envelope(s)` on a clean exit 0 with no captured output. A probe
measured that clamp firing 5 times in 613,775 calls under load
and never on an idle machine. The scenario now starts the job
again, which re-reads the clock away from the boundary, and
throws if it still refuses, so a recurrence fails loudly instead
of silently.

Separately, each check-in is its own HTTP request, so the order
they reach the mock server is not the order the SDK sent them in.
Under load the `ok` check-in overtook `in_progress`. This is
fallout from #20661, which moved these envelopes off stdout,
where they were inherently ordered. The assertions are now
`.unordered()`.

Failed roughly 1 run in 30 under load before, 120/120 after.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>