fix: implement missing GLSL.std.450 opcodes (#334)

[lkmavi]

Summary

• Fix incorrect Khronos opcode numbers (Determinant/MatrixInverse were 76/77; correct values are 33/34) and implement the remaining GLSL.std.450 ExtInst ops in the software SPIR-V interpreter.
• Add hal.Logger().Warn on unknown opcodes to stop silent zero corruption on CI/servers/software backend.
• Cover matrix (det/inverse), pack/unpack (incl. half + double), Modf/Frexp/Ldexp, bit-find, FaceForward/Refract, NaN-aware NMin/NMax/NClamp, Radians/Degrees/Fma; InterpolateAt* is identity + warn.

Closes #334

Test plan

• go test ./hal/software/shader — coverage 86.2% (≥86%)
• go test ./...
• golangci-lint run --timeout=5m
• CI green on this PR

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[kolkov]

Good work on filling the GLSL.std.450 gaps — the opcode renumbering from 76/77 to 33/34 is a critical fix (verified against GLSL.std.450.h), and the coverage of pack/unpack/matrix/split/bit-find is thorough.

Deep review found 2 critical bugs (empirically validated):

---

BUG 1 (Critical, VALIDATED): inverse3 and inverse4 missing cofactor transpose

glsl_matrix.go — the adjugate matrix is the transpose of the cofactor matrix. The output loop reads cof[0][c], cof[1][c], cof[2][c] (column c of cofactor), but should read cof[c][0], cof[c][1], cof[c][2] (row c of cofactor = column c of adjugate).

Empirical proof: for the non-symmetric matrix [[1,2,3],[0,1,4],[5,6,0]] (column-major), A * inv(A) produces garbage with the current code:

 27.0000 -22.0000   6.0000
 38.0000 -31.0000   8.0000
-12.0000  10.0000  -1.0000

After transposing the indices, it produces identity. Same confirmed for 4x4.

Tests pass because they only test diagonal/identity matrices where A^T = A — the transpose bug is invisible.

inverse2 is correct (hardcoded adjugate formula, no loop).

Fix for inverse3:

for c := 0; c < 3; c++ {
    out[c] = ValVec3(cof[c][0]*inv, cof[c][1]*inv, cof[c][2]*inv)
}

Fix for inverse4:

for c := 0; c < 4; c++ {
    out[c] = ValVec4(cof[c][0]*inv, cof[c][1]*inv, cof[c][2]*inv, cof[c][3]*inv)
}

Add a non-symmetric test case to prevent regression.

---

BUG 2 (Critical, VALIDATED): float32ToFloat16bits denorm path zeroes all denorms

glsl_pack.go, float32ToFloat16bits, case exp < 113 branch:

mant |= 0x800000                             // 24 bits
shift := uint32(126 - exp)                    // shift ∈ [14,23]
mant = (mant + (1 << (shift - 1))) >> shift   // now [1..10] bits
return sign | uint16(mant>>13)                // BUG: >>13 kills result

After >> shift, mantissa is already in 10-bit range. The extra >> 13 (correct for the normal path's 23→10 bit reduction) pushes all denorms to zero.

Empirical proof: exhaustive test of 6 denorm inputs — all produce 0x0000 with original, all produce correct bits with uint16(mant):

input           ORIG       FIXED      expected
3.051758e-05    0x0000     0x0200     0x0200
3.814697e-06    0x0000     0x0040     0x0040
5.960464e-08    0x0000     0x0001     0x0001
4.577637e-05    0x0000     0x0300     0x0300

Fix:

return sign | uint16(mant)

---

NOT A BUG (reviewed and dismissed):

float16bitsToFloat32 denorm uint16 wrap — initially looked like a bug (exp-- wraps uint16 from 0 to 65535), but exhaustive test of all 1023 f16 denorms shows zero mismatches between uint16 and int versions. The uint32 overflow in fexp << 23 accidentally produces correct IEEE-754 bits due to modular arithmetic. Not a bug, though using int would be cleaner.

---

Minor (not blocking):

• findILsb/findSMsb/findUMsb only handle scalar operands. SPIR-V GLSL.std.450 spec says these accept "scalar or vector of integer type" and return component-wise results. Fine for current usage, but worth a TODO.

---

Summary: Fix BUG 1 (inverse transpose) + BUG 2 (pack denorm >>13) + add non-symmetric matrix test. Everything else is verified correct — opcode numbers match Khronos header, NaN-aware min/max matches spec, snorm clamping correct, refract formula correct.

[lkmavi]

@kolkov Fixed comments. Round 2 Ready

[kolkov]

All 3 fixes verified locally: inverse3/4 transpose, f16 pack denorm, f16 unpack int type. Non-symmetric matrix tests + denorm roundtrip tests added. 22/22 packages pass, 0 lint issues.