Reject raw-encode domains that collapse after f32 conversion

[Alek99]

Review fixes for #187's _raw_wire_ok guard, split out of #221 when that PR retargeted to main (these fixes belong to the raw-encode code, which lives on this stack).

• P1 — the guard accepted domains whose f64 span survives an f32 cast but whose endpoints collapse to the same f32 ([1e30, 1e30 + 1e20]): every shipped value rounds to f32(lo) and the shader maps them all to the domain floor. It also accepted spans whose f32 reciprocal overflows ([0, 1e-45]), sending Inf into the map uniform. Both now require float32(lo) < float32(hi) and a finite f32 inverse span, falling back to the f64-normalized unit encode; tests pin both cases.
• P2 — the reciprocal bound is compared in f64 against np.finfo(np.float32).max, never cast through np.float32: the demoting cast warns on overflow for exactly the subnormal spans being rejected, and np.seterr(over="raise") setups would turn that into a FloatingPointError inside build_payload. The overflow test runs its probe under np.errstate(over="raise") to pin it.

Full tests/test_scatter.py passes (82); the fallback tests also pass with -W error::RuntimeWarning (no warnings emitted).

[greptile-apps]

Greptile Summary

This PR tightens the raw continuous-channel encoding guard. The main changes are:

• Reject domains whose endpoints collapse after f32 conversion.
• Reject domains whose f32 inverse span would overflow.
• Add fallback tests that raise on NumPy overflow warnings.

Confidence Score: 4/5

The raw-wire guard needs to reject domains whose mapping subtraction overflows in the shader.

Wide opposite-sign domains can pass the Python guard, and the same domains produce non-finite f32 arithmetic in the shader. The endpoint-collapse and reciprocal-overflow cases are otherwise covered.

python/xy/channels.py

T-Rex Logs

What T-Rex did

• T-Rex ran the requested verification, but its local artifact references were not uploaded.
• Before, the inverse-span regression in build_payload failed with a FloatingPointError: overflow encountered in cast.
• After, the strict-warning selection passed both fallback cases without warnings or exceptions, and full tests/test_scatter.py validation completed in 0.58 seconds.

_{Ran code and verified through T-Rex}

Important Files Changed

Filename	Overview
python/xy/channels.py	Adds endpoint and reciprocal checks, but still permits domains whose subtraction overflows in the f32 shader.
tests/test_scatter.py	Adds focused tests for endpoint collapse and inverse-span overflow without covering wide opposite-sign domains.

_{Reviews (1): Last reviewed commit: "Compare the f32 reciprocal bound in f64,..." | Re-trigger Greptile}

[greptile-apps]

Shader Subtraction Can Overflow

A domain such as (-3e38, 3e38) passes this guard because both endpoints are finite f32 values and the reciprocal is checked from a Python-f64 difference. The shader then evaluates a_cval - lo in f32, where values near the upper endpoint overflow to Inf, producing an invalid color coordinate or a saturated size. Reject domains whose f32 mapping subtraction is not finite.

Suggested change

	lo32, hi32 = float(np.float32(lo)), float(np.float32(hi))
	return lo32 < hi32 and 1.0 / (hi32 - lo32) <= float(np.finfo(np.float32).max)
	lo32, hi32 = float(np.float32(lo)), float(np.float32(hi))
	span32 = hi32 - lo32
	return lo32 < hi32 and span32 <= float(np.finfo(np.float32).max) and 1.0 / span32 <= float(
	np.finfo(np.float32).max
	)

_{Ran code and verified through T-Rex}

[Alek99]

Folding back per review direction — keeping #221 as the single PR. These two fixes patch _raw_wire_ok, which only exists on the #186–188 stack (not on main, where #221 now bases), so they cannot ride in #221's diff. They remain on branch alek/raw-wire-f32-guard (commits 1daca36, acc16f4), ready to cherry-pick into the stack or onto main when the stack rolls up.

[codspeed-hq]

Merging this PR will not alter performance

✅ 97 untouched benchmarks
⏩ 1 skipped benchmark¹

---

_{Comparing alek/raw-wire-f32-guard (acc16f4) with feat/append-skip-unchanged (fdd7ffc)²}

Footnotes

1. 1 benchmark was skipped, so the baseline result was used instead. If it was deleted from the codebase, click here and archive it to remove it from the performance reports. ↩

2. No successful run was found on feat/append-skip-unchanged (e7d8a57) during the generation of this report, so 2c62821 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩