Measure trycua/cua oracle gateability on this tree (cua-s1-forms port)

[lu-zero]

PR #3265 landed the cua-s1-forms port (TinyTransformerScorer, 706K params, f32) with .agents/oracles/cua.md pinned at trycua/cua 9bbfa7dd3e27ca7f1861ede70aaca390174493f9 and gateable = no, but the oracle has never been RUN against this tree — the pin was verified via the GitHub API only, so it carries no gate authority (AGENTS.md: an oracle is only gateable once it demonstrably builds and runs the model).

This issue owns that measurement and is the #N the oracle file's evidence field must cite:

1. Build/run the Laya-Python-equivalent reference path: execute the pinned trycua/cua TinyTransformerScorer + ByteCollator inference on the published cua-s1-forms checkpoint (2.8 MB, f32) and record outputs on a fixed workload.
2. Run our CuaS1ScoreInference on the same checkpoint and workload; compare probabilities to the oracle's recorded tolerance (spec: 1e-5).
3. Record the result in .agents/oracles/cua.md (gateable yes/no + evidence path) and in the row's records.

Local record: ISSUE-LOCAL-01M32WZMKVRBP4PH21KPHEDS3V (row MODEL-CUA-S1-FORMS).

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:regolo/glm5.2 [maki]