Numbers in any compare by value, whatever types they were boxed with

[ASDAlexander77]

Problem

=== and !== between two any values compared the boxes' bytes. When the two boxes had different sizes, e.g. a number and an s32, both answered false:

const g: any = 2.5; const h: any = 1;
g !== h   // false (should be true)
const e: any = 1.0; const f: any = 1;
e === f   // false (should be true)

DefaultLib's tests/math.ts and tests/shift_ops.ts check results with isEq(x: any, y: any) { if (x !== y) fail(); }, so a computed double compared with an integer literal could never fail. That is how isEq(Math.max(-1, -7), -1) passed while Math.max returned 2.2e-308 (docs/bugs/ntg_math_max_negative.ts, fixed separately in DefaultLib).

Also wrong, in the same code: loose == between two numbers with different tags (1.0 == 1 was false), and </> between numeric anys, which compared bytes.

Fix

AnyCompareOpLowering: when both tags are of the numeric kind, both values are read as f64 and compared with the operator's float predicate (oeq, une, olt, ...). The tag's first letter says float (f, number), unsigned (u) or signed, and the box's data size says the width. The existing number-unboxing helper only knew number, s64 and s32; it is now this shared member. For other kinds whose sizes differ, !== now answers true, as != already did.

Testing

• New 00any_number_compare.ts (compile + JIT) covers mixed widths, s64/u8/f32/s16, ordering, NaN, and non-numeric kinds. It fails on the current release (assertion failed: 2.5 !== 1).
• Full ctest: 3915/3915 pass.
• DefaultLib's tests/math.ts run with this compiler now reports fail: 2.22507e-308 !== -1 (the real Math.max bug). With that line removed, the rest of math.ts passes, and so does shift_ops.ts, so nothing else was being hidden in those two files.

🤖 Generated with Claude Code