[tannevaled]

Closes #820.

FFT.fft()                 # before: ArgumentError: fft: runtime error: index out of range [0] with length 0
                          # after:  ArgumentError: wrong number of arguments (given 0, expected 1)
Image.new(2, 2).set       # after:  ArgumentError: wrong number of arguments (given 0, expected 5..6)

All 17 of FFT's methods and 14 of Image's indexed args[0..3] with no count check, so a missing argument surfaced as a Go panic, recovered and relabelled. Its class was accidental too — ArgumentError only because the recovery path labels every panic that way. NDArray was already clean because it uses ndArgc, which both now use.

The counts live in the signature, not in the body

def(name, min, max, fn) for FFT, and d / s for Image's instance and module methods. A check written inside a body can be forgotten by the next method someone adds; a parameter cannot. The three wrappers that forward to d — encode, unary, scalar — pass their own fixed arities, so the sixteen unary filters and the two scalar ones are covered by one line each instead of eighteen.

Each arity comes from reading the body, and the ranges are the optional arguments a caller really has:

fftfreq / rfftfreq / psd	1..2 — numpy's sample spacing d
spectrogram	4..5
Image#set	5..6 — the alpha that defaults to 255
everything else	fixed: 0, 1, 2, 3 or 4

The test is a census, and says how far it walked

names.sort.each { |m| bad << m if (recv.send(m) rescue $!.message.include?("runtime error")) }

A census covers a method added later without it being registered anywhere — which a list of names does not. And it asserts the number walked (17 3 34) as well as the empty result, because a census that reads nothing otherwise looks exactly like a clean module. #save is held out, since it writes a file, and checked by name instead.

Nineteen message rows pin MRI's wording for each shape: fixed 1/2/3/4, both range forms, and the too-many-arguments form. The count is only half of it — someone rescuing ArgumentError and printing e.message has to be told what they passed wrongly, and "index out of range" does not tell them.

Controls

Eleven answers that were right before and have to stay so, since every method is now wrapped: FFT.fft([1,2,3,4]) → [10, -2, -2, -2], fftfreq with and without spacing, hann(3), the rfft/irfft round trip, Image#set/#get with and without alpha, the filters' shapes, #to_png, and #to_s — which is aliased from #inspect's native function and so inherits whatever wrapper #inspect got.

Four negative controls, each verified to have really changed the file:

break	result
drop the check in FFT's def	RED
drop the check in Image's d	RED
take fftfreq's optional spacing away	RED
take Image#set's optional alpha away	RED

The existing Image, FFT and canvas tests pass unchanged; gofmt, go vet and credscan clean. Branched from main so it is independent of #819, which found the defect but does not fix it.

Measured 2026-10-18 with go1.27.2, CGO_ENABLED=0.

🤖 Generated with Claude Code