[AskAlexSharov]

#32 removed NOSPLIT from the amd64 kernel so the runtime regains a preemption point in the absorb and squeeze loops. Nothing stops it coming back, and the failure is invisible in throughput benchmarks — it only shows up as a stop-the-world pause in whatever else the process is doing.

TestKernelKeepsStackCheck

Reads the TEXT directives and rejects both ways of losing the check:

• NOSPLIT, which drops it whatever the frame size;
• a frame under 128 bytes, because the assembler marks a leaf that small NOSPLIT on its own (abi.StackSmall, arm64).

The second one is the trap: deleting the keyword alone changes nothing if the frame is small, which is why #12 would have reintroduced the stall on arm64. The test reads both .s files regardless of the runner's architecture, so an arm64 CI job still catches an amd64 regression.

TestSTWPauseWhileHashing

Measures the property the directives only imply, using the runtime's own /sched/pauses/stopping/gc:seconds histogram while hashing 64 MiB with GC pressure on another goroutine. It reads the histogram delta rather than the process-lifetime max, and waits before reading so a blocked stop-the-world is recorded first.

Verification

Both tests pass on master. Putting NOSPLIT back on the kernel of each architecture:

	now	with NOSPLIT restored
amd64, EPYC 4344P	20.48 µs	83.886 ms
arm64, M4 Max	114.69 µs	58.720 ms

The budget is 5 ms — two orders of magnitude away from both the healthy and the broken value, so there is nothing to tune. The test takes 1.3 s and skips under -short, when GOMAXPROCS < 2, and when the CPU has no assembly backend (the pure-Go path is preemptible, so there is nothing to measure).

[AskAlexSharov]

Merged into #35. The test is red on master because keccakF1600BMI2 is still NOSPLIT there, so it only goes green together with that fix — keeping them in one PR avoids landing a knowingly-failing test. Closing in favour of #35.