[tannevaled]

The updates of Renovate's #106, applied on current main with every module tidied:

• Root module:
  • go-fft/fft v0.28.11 → v0.29.0
  • go-simd/floats v0.15.0 → v0.17.0
  • the go line moves to 1.27.3
• internal/kernels/asmgen: go-asmgen v0.23.0 → v0.23.2. go generate rewrites sum_*.s with no change.
• bench: go-scikit/refbench v0.1.2 → v0.3.0, and its indirect requirements follow the root module.

#106 failed in "perf (refbench CI mode)" with go: updates to go.mod needed; go mod tidy. bench/ reaches the root module through replace ../, so its indirect requirements have to follow every root bump, and Renovate's run does not re-tidy it after the root module changes.

Also, after review:

• fft.Options gains Exact and FMA. They are passed to go-fft v0.29.0's options of the same names. Exact is go-fft's reference arithmetic, today its default. FMA fuses the twiddles on amd64 with FMA3. Asking for both is refused with ErrArgs. TestArithmeticOptions covers this: Exact gives the default bits, FMA stays within 64·eps·max|X| of the default, both together are refused, and the repeated-axis path passes both through. fft stays at 100 % coverage.
• README: the accuracy contract's Exact line now names fft.Options{Exact: true}, since the README promised that option.
• internal/kernels/asmgen/arm64/gen.go: the comment still said a SIMD max was out of reach because of bit identity, but Max and Min have come from go-simd/floats' NEON kernels for a while. paritycheck repo (fleettools v0.106) refused that sentence; it passes again. The .s files are unchanged.

Local runs: go test ./... passes (root, fft, kernels, linalg, random), and go vet ./... in bench is clean.

Supersedes Renovate's pull request: Refs #106 (Renovate closes it once main has these versions).

🤖 Generated with Claude Code