[enesismail]

What this changes

The npm wrapper's row "the waits between attempts are one second and then two" timed the pauses between three
download attempts at the server, and required each one inside a window: 900-1800 ms, then 1900-2800 ms. Once, a hosted
Windows runner stalled for several seconds. The second pause measured 5407 ms and the row went red, with the schedule
correct. In the same job, an unrelated proxy row took about forty times its usual time. A stopwatch on a shared
machine measures the machine.

The schedule is now read from the child process:

• A preload in the test tree, npm/test/helpers/recorded-waits.js, beside the two that are already there. It
  replaces the global setTimeout in the installer's process and writes down every delay asked for, and when each one
  fires. Each line is written synchronously, because the installer exits the moment its last attempt fails. Every call
  goes to the real timer with the same delay, so each pause is still spent. The file it writes to is named by a
  variable only the preload reads.
• The row requires the whole record:
  set 30000, set 1000, fired 1000, set 30000, set 2000, fired 2000, set 30000. It also shows that each attempt arms a
  fresh deadline, and that each pause fires before the next attempt starts.
• The server still counts three attempts, and the pauses it sees keep their lower bounds, 900 ms and 1900 ms.
  • The ceilings are gone. The server stamps an arrival and drops it in one step, so a stalled machine can only lengthen
    a pause.
  • The floors catch a pause that was recorded and never spent.
  • Arrivals are now stamped with performance.now(). The other users of that listener only count arrivals.
• The row is three subtests over one install. It prints its record and both pauses before asserting anything, so
  every log carries a sample, including a red one.
• Comments: the file header and the pin row's comment now say what is taken where. The pin claims only that each
  attempt arms a fresh deadline, not that the deadline ends one.

npm/install.js is unchanged.

Mutations

Each was made in the working tree, run, and restored by checksum:

mutation	count	record	floors	other rows
a pause after the last attempt (< to <=)	pass	fail	pass	the row before this change stays green under it
second pause made one second	pass	fail	fail	the pin stays green
a fourth attempt with a third pause	fail	fail	pass	the pin and three other rows fail
await dropped before the pause	pass	fail (no fired lines)	fail
the preload fires each pause at once, recording it as asked	pass	pass	fail

A dropped await fails the record by order, however loaded the machine is. A preload that fires the pauses at once is
caught by the floors alone. That is normally certain, but a long enough stall could in principle lift a near-zero
pause above a floor.

Measured

Before the change, the row's recorded results across every retained CI run: 2,575 readings on the five jobs that run it,
one red. On the Windows job, fifteen passes took over 3.8 s, the slowest 7.5 s.

Locally, before the change, 500 runs idle and 500 with every core busy: no red, and no pause short of its length.