[Docs] Require cold-load timing for in-process backends

CHANGELOG.md

@@ -59,6 +59,10 @@ The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); ver
 
 ### Changed
 
+- Backend contribution guidance requires cold in-process model load timing, profiling above 5 s, and
+  complete checkpoint coverage before skipping random weight initialization.
+- `--model laya` loads in about 3 s instead of about 35 s: the encoder is built with transformers' weight init
+  off, since the checkpoint replaces every weight. Weights and answers are unchanged.
 - Important agent/inference PRs require an application + System1-Agents + System1-Omni video, following
   PR #35's worked example. Contributor skills, recipes and the PR template retain missing demos as review gaps.
 - `--model laya` no longer draws the encoder's random weights before the checkpoint replaces them, which took most

docs/decision-models.md

@@ -96,6 +96,12 @@ real servers in `tests/data/served_laya/`. Nothing patches `httpx`.
 3. `tests/test_decision_models_<backend>.py` with `Test<Backend>Contract(DecisionModelContract, IsolatedAsyncioTestCase)`
    plus the backend's mapping tests; a fake for its SDK lives in that file.
 4. An optional extra in `pyproject.toml` and an env block in `.env.example` when it needs a dependency.
+5. For an in-process backend, time a cold `from_env()` in a fresh process and report the load time in the PR,
+   separately from warm inference and episode timing. Profile loads above 5 s. If the loader builds a model
+   from its config (`AutoModel.from_config` or a bare `nn.Module`) and then replaces every weight with a
+   checkpoint, it can skip random initialization with transformers' `no_init_weights`, but only if loading
+   fails on missing keys (`strict=True` or an equivalent check). Prefer fixing this in the model's package;
+   Laya added the skip in 0.3.9. This step does not apply to backends served over HTTP.
 
 ### Laya browser input shaping