Several things.
At once.

01 / schedule

Schedule almost nothing.

Creating, joining and collecting zero-work tasks. Request/response serialization is included; these are not engine-internal scheduler nanoseconds.

02 / tiny

A little work in every branch.

Eight exact recurrence iterations per job. Large task sets are expensive stress cells with five samples for every participant.

03 / fan

Fan out. Collect everything.

Every branch has a unique ID. The oracle checks exact values, missing work and duplicate results. This is the same flat create/join shape as tiny arithmetic, retained as a separately sampled fan-out cohort; it does not imply a different executor capability.

04 / wait

Wait together.

Nift blocking sleeps occupy native async workers. Event loops suspend promises/tasks and Go suspends goroutines. Equivalent waits do not imply parallel computation. On this four-worker native pool, 1000 × 10ms blocking sleeps have a 2500ms work/worker lower bound; suspended timers have a 10ms critical-path floor before scheduling and result overhead.

05 / cpu

Use all four cores.

The same 64 × 200000-iteration kernel at one, two and four workers. Python uses ordinary-GIL multiprocessing; Node/Bun use worker isolates. Pools are created inside each request.

Four-worker efficiency: Nift 95.2%, Python 77.5%, Node 54.5%, Bun 67.4%, Go 85.9%. Ratios of measured medians, without a significance claim.

06 / pool

Give the queue a limit.

One, two or four workers dispatch a fixed 64-job set. Light, medium and heavy bodies reveal the tradeoff between pool setup and useful computation.

07 / queue

Talk to each other.

Capacity 64. Go uses channels, Python queue.Queue, JS an authored cooperative promise queue. Nift uses a mutex-protected ring and 1ms empty/full backoff; this is a documented channel/condition-variable gap.

08 / sync

Share state carefully.

Genuine shared atomics only. Python atomic cells and JS native-mutex cells are unavailable (—). Mutex critical sections contain one or eight increments; all counts are exact.

09 / process

Launch. Capture. Join.

Four concurrent argv-launched identical helpers. No shell startup. Every stdout value and expected exit status is checked, including failures.

10 / files

Read together.

Pregenerated 128-byte files, exact byte hashes, four concurrent operations. This is cached runtime/file orchestration, not a disk benchmark.

11 / mixed

Different work, one result.

Mixed jobs combine arithmetic, a wait and a child process. Nested parents create four children. Failures collect all outcomes and let successful siblings finish.

12 / memory

Keep work alive.

All tasks are staged behind a gate while tree RSS is sampled. Final Nift10k held tasks are outside the certified matrix: a separate probe timed out before READY at45seconds with sampled peak4,137.840MiB below its6GiBguard. This time-censored diagnostic is not a plotted wall/memory result or proof of8GiBinfeasibility. Queued Nift futures need not have started running. Shared pages may be double-counted; this is not unique physical memory.

Unranked5ms sampled peak accounting, separate from post-READY held checkpoints. Pre-exec supervisor floors and shared mappings can distort it; it cannot rank memory efficiency. The final noop control rose16.75→80.75MiB for0→64MiBparent allocation; no subtraction.