fix(config): reject doubled provider prefixes in agent model ids

[w7-mgfcode]

Summary

Runtime config accepted a malformed agent model id with a doubled provider prefix, which then produced a 404 from the Gemini API at agent-run time.

Details

• PATCH /config/ai persisted agent_default_model = google-gla:google-gla:gemini-3-flash-preview (returned 200).
• PydanticAI treated the first google-gla: as the provider and sent google-gla:gemini-3-flash-preview as the model name.
• Google returned: 404 NOT_FOUND — models/google-gla:gemini-3-flash-preview is not found for API version v1beta, or is not supported for generateContent.
• Net effect: the primary model is unusable until the id is hand-corrected; the failure only surfaces when the agent is invoked (not at config-save time).

Expected

Config validation should reject or normalize a doubled provider prefix in agent_default_model / agent_fallback_model:

• Reject (422) ids where the provider prefix appears more than once (e.g. google-gla:google-gla:...), or
• Normalize by collapsing a duplicated leading provider prefix.
• Apply to all supported providers (google-gla, google-vertex, anthropic, openai).

Tests

• Schema-level: validator rejects/normalizes doubled-prefix ids.
• Service-level: PATCH /config/ai with a doubled prefix returns 422 (or stores the normalized value).
• Route-level: happy path + the doubled-prefix error path.

Context

Found while investigating an experiment-agent stream failure ("All models from FallbackModel failed"). The doubled prefix was the primary-model leg of that failure. Companion issue covers surfacing fallback sub-errors to the UI.