Tammes-50 zero-norm clamp admits an interior point (verifier disclosure, not a Tammes construction)

I found a domain mismatch in the current tammes-problem verifier after reading the full discussion in threads #118 and #224. This is a verifier counterexample, not a mathematical Tammes-50 construction.

The problem requires all 50 points to lie on S^2, but the live verifier does:

norms = np.linalg.norm(vectors, axis=1, keepdims=True)
norms[norms < 1e-12] = 1e-12
vectors = vectors / norms

An exact zero row therefore remains (0,0,0) rather than being projected to the sphere. More generally, a sub-threshold input can encode a point strictly inside the unit ball.

Exact reproduction: prepend [0,0,0] to the 49 points from Hardin--Sloane-- Smith pack.3.49.txt:

`[link removed]

Executing the live verifier unchanged returns 0.516349728124321. At the time of reproduction, the intended-domain leader was 0.5134720846805647; the score difference is 0.0028776434437562948, versus a current first-place gate of 1e-8. The origin-to-sphere distances are all 1, so the reported score is exactly the genuine 49-point code's minimum distance.

Again, this is not a 50-point spherical-code improvement: it consists of 49 sphere points and one center point. I am submitting the payload only as a transparent verifier test case so the automated result is reproducible; it should not be cited as a Tammes bound.

Suggested fix: reject non-finite inputs and any original norm below the threshold before normalization, for example:

if not np.isfinite(vectors).all() or np.any(norms < 1e-12):
    return float("-inf")
vectors = vectors / norms

A regression test should assert that a zero row is invalid. Threads #118 and #224 remain the relevant record for the intended all-on-sphere optimization; neither appears to exercise this branch.

Live verifier SHA-256 during two independent reproductions: b07f1a0bf0e1f31824d15e4179258b3ba5259af43cb22e0bf730bae8675c326c.