[Know-It-All-Brother]

Part of the Ampere accuracy series (#2040). It depends on #2041 for the test: the commits before this PR's own are #2041's and drop out once it lands, so this PR's own change is the last commit.

What and why

.NE is the ordered not-equal: it's false when either operand is NaN (.NEU is true then). The translator emits SPIR-V's OpFOrdNotEqual, which is right, but spirv-cross writes it as MSL's !=, which is unordered. So FSETP.NE held for a NaN operand, and a NaN guard let the NaN through. ptxas also builds setp.equ from the negated FSETP.NE, so that was wrong too.

The obvious rewrite, (a < b) || (a > b), doesn't work: Metal's fast math assumes no NaN and folds it straight back to !=. I checked that on the GPU. GetFPComparison now writes a != b && !isnan(a) && !isnan(b), and Metal keeps isnan under fast math, as the unordered comparisons already rely on.

The helper is shared, so Maxwell's FSETP and FSET get the fix too.

Reference entry: "FSETP.NE and FSET.NE held with a NaN operand" in docs/reference/shaders.md.

How to verify

DMSH_TEST_XLATE=dm/gfx/tools/dmsh_xlate_bin/dmsh_xlate AMPERE_ORIN_ONLY=fsetp,fsetp_ftz,fset_bf,fset_u32,fsel_cmp python3 dm/gfx/tests/test_dmsh_ampere_orin.py

Op	main	this PR
fsetp (14 compares)	66 of 14,336 lanes differ, 47 of them with a NaN operand	16, all f32 denormal inputs (Metal flushes them)
fsetp_ftz	45 of 4,096	0

On the full 16,384-case run, 800 cases of each differed before, every one with a NaN operand. The test takes fsetp and fsetp_ftz out of KNOWN, so it fails on main's translator and passes here.

Checked

• Tests: test_dmsh_ampere_orin.py, and every other test_dmsh_* gives the same result as on main (test_dmsh_ampere_mem.py, test_dmsh_meshstage.py and test_dmsh_pairs.py fail on main the same way). test_nsp_chain.py, test_xlate_anycpu.py, test_docs.py.
• Translator: the dlls are rebuilt from this tree with .NET 10, tool_lock.py --update-lock and dmsh_chain.py --header were run (chain 230bb40d4e58).
• Devices and OS: MacBook Pro M5 Max, macOS 27.
• regress.py: not run. My only local Switch 2 build's shader package comes from an older shader split than main's map, so main's chain can't build it. I'll add a report once I have a fresh build. A scan of Ocarina of Time's shader code finds only a handful of FSETP.NE.
• Games run: none. A scan of Ocarina of Time's shader code finds only a handful of FSETP.NE, so I don't expect a visible change there.

Docs

• docs/reference/shaders.md has the entry, and python3 dm/tools/docs_symptoms.py was run
• A chain tool changed: dmsh_chain.py --header was run, and dmsh_chain.h, the translator dlls and the lock are part of the change
• No game files, keys or data extracted from a game