fix(capy): label config-sourced agent_model, note the assumption, skip it inside eval runs (GST-982 review)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

docs/capy-install.md

@@ -473,7 +473,8 @@ make <dir> writable, or set GSTACK_STATE_ROOT to a writable directory
 and removes the file an earlier passing run left beside the verdict it rewrites
 (a profile check that keeps a fuller ok verdict leaves that verdict's file).
 With `--model <id>`, `GSTACK_MODEL=<id>` in its environment, or else an
-`agent_model` in the session's `GSTACK_PREFERENCES_FILE` (GST-982), an ok run other than `--for review-voice` also prints `LOADOUT: …`
+`agent_model` in the session's `GSTACK_PREFERENCES_FILE` outside an eval run (GST-982, with a
+`LOADOUT_ASSUMED:` note), an ok run other than `--for review-voice` also prints `LOADOUT: …`
 and `SKILL_ROOT: <dir>` just before `GSTACK_LAUNCH`. For a loadout other than
 `full`, `GSTACK_LAUNCH_FILE` then names `<verdict file without .json>.loadout-<name>-<digest>.env`,
 a copy of the verdict's launch file that adds `GSTACK_LOADOUT_SKILL_ROOT`; the

docs/loadouts.md

@@ -55,7 +55,7 @@ A model id is lowercased, its `provider/` prefix and a trailing `[…]` are drop
 
 ### Capy
 
-A plain `bash scripts/install.sh --check --for <skill>` picks the loadout from the owner's `agent_model` preference, so the skill needs no model sentence (`source=preferences` on the line below). A session can still name its own model with `GSTACK_MODEL=<id>`, which wins over the preference and, unlike `--model`, is safe on a checkout whose installer predates loadouts: that installer ignores it instead of refusing an unknown argument. The installer prints two lines before the launch lines:
+A plain `bash scripts/install.sh --check --for <skill>` picks the loadout from the owner's `agent_model` preference, so the skill needs no model sentence (`source=preferences` on the line below, or `source=config` when this machine's `config.yaml` sets `agent_model`). The preference names the model Garry starts sessions on, not the model reading the skill, so the installer says so in a `LOADOUT_ASSUMED:` line, and it skips the preference inside an eval run (`GSTACK_EVAL_DIR` or `EVALS_RUN_ID` set), where the model under test may differ. A session on another model names its own with `GSTACK_MODEL=<id>`, which wins over the preference and, unlike `--model`, is safe on a checkout whose installer predates loadouts: that installer ignores it instead of refusing an unknown argument. The installer prints two lines before the launch lines:
 
 ```
 LOADOUT: lite model=claude-opus-5-5 source=GSTACK_MODEL digest=3f2a9c01d4e7 reused=yes