fix(web): ship placeholder judge model so a clean clone resolves & runs

[kongyoongkeong-lab]

Summary

The Web benchmark (configs/bench/wb-bench-web-v1.0.yaml) pins llm_judge.model: kimi-k2.7-think by default, but configs/models/** is gitignored (only _template.model.yaml ships). On a clean clone, resolve_manifest therefore fails every Web run with:

llm_judge.model='kimi-k2.7-think' does not resolve to a model config
at .../configs/models/kimi-k2.7-think.yaml

All 70 Web tasks contain llm/vlm judge items, and the in-container CompositeVerifier raises when the verifier LLM route is not configured (route_ready=False), so the Web track is currently unrunnable out of the box.

Fix

• Un-ignore and ship configs/models/kimi-k2.7-think.yaml as an env-only placeholder (mirrors _template.model.yaml). Credentials still live in .env (KIMI_BASE_URL / KIMI_API_KEY); the file names only the env vars to read.
• The web bench default (enabled: true, model: kimi-k2.7-think) is left intact — this only restores the model file the default points at, so the LLM/VLM judge route resolves and the verifier stops raising.

Users with a Moonshot/Kimi backend fill the two .env vars and run as before. The slug now resolves cleanly; a missing KIMI_BASE_URL surfaces a clear runtime error instead of a silent manifest-resolution failure.

Test plan

• resolve_manifest no longer reports the missing-model error for a Web job (model slug now resolves to a config).
• Placeholder kimi-k2.7-think.yaml tracks (gitignore exception added).
• Maintainer: confirm the placeholder shape matches the intended Kimi/OpenAI-compatible endpoint (no key/credential shipped).

Notes / alternatives considered

• Simply changing the web default to <model-slug> + enabled: false passes the resolver but breaks scoring: the verifier raises on the 70 tasks that require an LLM/VLM route. Keeping enabled: true and restoring the model file is the minimal change that keeps the Web track actually runnable.
• If the project would rather not ship a Kimi-specific file, an alternative is to document "Web requires a local kimi-k2.7-think.yaml (cp from _template)" in the README and relax the gitignore similarly — happy to adjust.