docs(models): record full glm-5.3 checkpoint screening

[earlvanze]

Scope

Record the full LibertAIDAI/GLM-5.3-NVFP4 checkpoint on the existing glm_moe_dsa path, preserving GLM-5.2 aliases, with config regression coverage and measured Vast screening evidence. This is full GLM-5.3, not Flash.

Measured evidence

One RTX PRO 6000 Blackwell Max-Q 96 GB, driver 595.84, Ubuntu 24.04.3, 36.48 effective CPU cores and a 1.87 TiB container memory limit. Runtime commit 55120d9. The report includes the exact inspected launch command and immutable links to all raw request results.

Initial 32-token direct-worker screen: c8 3.937 aggregate output tok/s; c16 offered against eight running slots 6.297. All 24 requests truncated during reasoning; zero final coding answers. Neither result qualifies the requested c8 target. A separate arithmetic smoke returned the correct answer but exposed a reasoning-parser formatting issue.

Verification and remaining work

Locally ran server tests plus the new model-config tests: 541 passed. That run included a separately prepared, uncommitted parser fix; this PR does not include that fix and does not claim patched-runtime validation.

Draft: longer repeated benchmarks, final-answer checks, actual serverless routing and idle/cached-restart lifecycle remain pending. No production-readiness or per-user 15 tok/s claim. No performance-engine change or A/B speedup claim.

AI-assisted under operator direction; preliminary evidence is intentionally retained, including failures.