Opt target 3 — pipeline: 1.18x Node, but 1.85x instructions at 1.05x cycles

This one is a different animal from the other two, and the difference is the whole ticket.

Measurement — quiet M1 mini, 5 shuffled interleaved repeats, verdict CLEAN, perry fefdc367b

Perry retires 85% more instructions but spends only 5% more cycles, at nearly double Node's
IPC. That combination says the extra instructions are cheap, well-predicted, and largely free —
so cutting the instruction count is not where the 18% comes from. The wall gap is somewhere a
sampling profiler cannot see it.

asyncpipe had exactly this shape and it has now been fixed — it sat at 1.07x wall against
1.71x instructions, and #8570 flipped it to 0.84x by suppressing redundant same-thread
promise-drain wakeups and removing per-element closure/guard/chained-promise allocations from
the intrinsic Promise.all path. Its instruction ratio fell 1.71x → 1.31x as a side effect.
That fix is the strongest available hint for this row, and #8243 has prior analysis.

What NOT to do

Do not reach for a sampler and start cutting instructions. asyncpipe taught this: aggregating
sample over 14 runs of an ~88 ms program yielded one usable self-time sample, because the
process is almost always parked rather than running. A profiler will show you a flat, uninformative
profile and you will conclude there is nothing to fix.

Method

Instrument the event loop and the scheduling machinery rather than the code:

• count and time event-loop wakeups, and how many are redundant with a drain already in flight
• count microtask-queue drains, and jobs consumed per drain
• count timer registrations and their latency from due-time to run
• look for per-element allocations on the intrinsic paths, as perf(runtime): eliminate redundant asyncpipe microtask work #8570 found in Promise.all

PROFILE_NOTIFY_DRAIN_SUPPRESSED_COUNT exists from #8570 and is a working example of making
this class of subject countable rather than merely observable.

Acceptance

• Wall time is the primary signal here, unusually — the instruction count is already
  known not to be the constraint. Report both, but do not claim a win on instructions alone.
• Because the bench host is contended and this row is ~100 ms, use many interleaved repeats and
  report spread; treat a wall delta as real only when the ranges are disjoint.
• All 20 corpus programs byte-exact; whole-corpus timing before and after, every row moving
  more than 1% reported.
• RSS is 0.25x Node. Do not trade it.