Match OpenSCAD's argument and arithmetic diagnostics

[revarbat]

Why

ord(undef) printed no warning where real OpenSCAD does. Chasing that found the same hole everywhere: we produced the right undef value for a bad argument or operand — silently — for every builtin and every arithmetic operator. A misspelled variable reaching abs(), len() or + said nothing at all.

How it was checked

Generated 1002 cases and ran them through OpenSCAD 2026.02.01 (~/Desktop/OpenSCAD-dev.app, not the 2022 build in /Applications — it differs) and through us, comparing both value and warning text:

Harness	Cases	Covers
operator matrix	392	every ordered pair of {undef, number, bool, string, vector, range, function} for + - * / % ^ and unary -
builtin conversions	205	every builtin with each argument wrong in turn
search()	225	needle × haystack × num_returns × index_col
undef survey	87	the original question, incl. all comparators
arity / edge / nested	93	missing + extra args, div-by-zero, ragged matrices

All 1002 now agree, except three rands() cases (random) and version() (the binary is 2026.02.01; we report 2026.01.01).

Diagnostics added

All reference-exact, none invented:

undefined operation (undefined + number)              + - * / % ^, unary -
NAME() parameter could not be converted: argument N: expected T, found T (v)
NAME() parameter could not be converted: vector element N: ...      max/min
NAME() number of parameters does not match: expected N, found M
vector*vector requires matching lengths (2 != 3)      + matrix siblings
Multiplication is undefined on empty vectors
Matrix must be rectangular / must contain only numbers. Problem at row N
Incorrect arguments to norm()
Invalid vector size of parameter for cross()
Invalid value in parameter vector for cross()
Invalid entry in search vector at index N, required ...
Parent module index (N) greater than the number of modules on the stack

Value fixes found along the way

Each verified against the binary, not inferred:

Expression	Was	Now (= OpenSCAD)
-7 % 3	2	-1
5 % 0	undef	nan
0 ^ -1	undef	inf
5 / [1,2]	undef	[5, 2.5]
[] * []	0	undef + warning
search(undef, "abc")	[]	undef
search("a", undef)	undef	[]
search("a", ["a","b"])	[0]	[] + warning
has_key(obj, undef)	false	undef
chr(65, 66)	"A"	"AB"
chr(-1), chr(1e9)	raw invalid UTF-8	""
echo(object(a=1))	object(a = 1)	{ a = 1; }
abs(1, 2)	1	undef + warning

⚠️ % is the one change that can alter rendered output. Ours was a floored/Python modulo; the reference is C's fmod, so the result takes the left operand's sign. A real BOSL2 model (spur_gear + prismoid + path_sweep) exports byte-identical before and after, so nothing in normal use depended on the old sign.

Also: oscTypeName called ranges and functions "undefined", which put a phantom undef operand into every message naming one.

Comparators were already correct and are untouched — < > <= >= warn and answer undef; == != never warn.

Note on search()

The port follows current upstream, which includes a crash fix (f2be6d0d2, 2026-07-02) that the 2026.02.01 binary predates — a non-string entry at index_col aborts that binary with bad_variant_access. Those shapes are excluded from the binary comparison and follow the fixed source.

Tests

878 pass under OSCAD_BYTECODE_VM=0 and =1. BelfrySCAD's own 794 pass against this build. Five existing tests encoded the old wrong behavior and were corrected with the reference output alongside.

Version 0.32.2 → 0.33.0 (minor: real behavior changes).

🤖 Generated with Claude Code