docs: the T4 has never run a kernel here

[vyncint]

docs/LIMITATIONS.md says both of these, twenty lines apart:

line 147: Only 8.6 (A10G) and 7.5 (T4) have ever had a kernel measured on them here.

line 167: All published numbers in this repository are from the A10G; nothing has been measured on a T4.

They cannot both be true, and the evidence in the repository is with the second:

• docs/research-baseline.md:16 — the tier-2 box is an AWS g5.xlarge, NVIDIA A10G (sm_86, cc 8.6), "chosen over the T4 by the operator; barely above T4 spot"
• model-calibration.toml:11 — device = "NVIDIA A10G (cc 8.6)", and it is the only device in the file
• runs/ holds one committed sweep, reduce-stable-metal, on an Apple M4 Pro

So line 147 claimed a hardware validation at cc 7.5 that never happened.

This is not a typo in a footnote. README.md points at this file twice — "the honest edges live in docs/LIMITATIONS.md. Read that before trusting a result" — and the paragraph exists specifically to tell a reader which rows of the model's device table are experience and which are only documented capacity. Someone deciding whether to act on a --cc 7.5 ranking was reading this exact sentence to make that decision, and it told them the part had been measured.

Five of the six rows are capacity rather than experience, not four. The corrected paragraph says so, names the two files that establish it, and keeps a parenthetical recording the contradiction — a claim that was wrong once is worth marking, so nobody restores it from memory.

I checked the rest of the file for the same shape: no other capability claim in it is contradicted by research-baseline.md or model-calibration.toml.

Closes #73