simd(s390x): vector max/min on z14+ (VFMAXDB Java mode)

[tannevaled]

What

• maxVXE/minVXE, generated by go-asmgen, using VFMAXDB/VFMINDB and the scalar WFMAXDB/WFMINDB in mode 1, Java's Math.max/min: a NaN operand gives NaN, and −0 orders below +0. That is NumPy's and Go's rule. Four two-lane accumulators from 8 elements up, then a fold and a scalar tail.
• Gating: AT_HWCAP VXRS_EXT (bit 13, vector-enhancements facility 1, z14+). On the z13 that Go still supports, the scalar reducers remain.

Tests on the real IBM z17 (ssh linuxone)

• TestMaxMinSIMD and TestMaxMinSIMDSpecialsInVectorLanes (±0, ±Inf in every lane) pass, with useVXE = true. TestVXEDetected re-runs both with the kernel forced off.
• The fence tests (both flags off in the fallback) and the parallel stress test pass, plus the full suites.
• Sabotage, caught: mode 0 (IEEE maxNum, which drops NaN) gives n=16 NaN@0: maxSIMD=1421.09…, want NaN.
• Coverage 100% with -race; go vet on 5 architectures.

Measured (z17, one core, hostload 0.1% busy, benchstat over 8 interleaved runs, p = 0.000)

	before	after
Max 1 Ki	2.58 µs	112 ns	−95.7%
Max 16 Ki	41.6 µs	1.78 µs	−95.7%
Max 4 Mi	10.7 ms	0.81 ms	−92.4%
MaxAxis(1)	2.69 ms	0.17 ms	−93.6%

Sum is untouched and serves as the control: ±0.6%.

Docs: README (max/min SIMD on s390x z14+), perf.md, plan, SECURITY.md.

Version: v0.9.1 (performance, no API change).

🤖 Generated with Claude Code