Keep DeepSeek out of public leaderboard until benchmark requests complete reliably

DeepSeek is configured only as an experimental/manual model path for now, not as a default benchmark model or public pending leaderboard entry.

Reason for exclusion from the public leaderboard:

• Credentials work, and a raw DeepSeek Flash toy JSON call succeeds.
• A real PolicyBench one-household, one-output deepseek-v4-flash smoke completed, but took about 109 seconds and used 8,207 reasoning tokens for one requested output.
• deepseek-v4-pro did not finish the same one-household, one-output smoke within 150 seconds.
• Earlier full multi-output PolicyBench attempts timed out or hung.

Current repo policy:

• DeepSeek remains in EXPERIMENTAL_MODELS so it can be manually probed if the API behavior improves.
• DeepSeek is excluded from default runs, app model metadata, and public pending-model messaging.

Acceptance criteria before adding it to the public leaderboard:

• Complete a representative multi-output smoke without manual repair or process intervention.
• Produce parseable JSON with nonempty explanations.
• Have latency and token usage that make a 100-household run practical.
• Pass the same analysis/rebuild path as other public leaderboard models.

Start by locating EXPERIMENTAL_MODELS, the default-run configuration, app model metadata, and public pending-model messaging. Run the representative multi-output PolicyBench smoke and inspect the analysis/rebuild path used by public models. Done means DeepSeek remains excluded until it produces parseable, explained outputs with practical latency and token usage without manual intervention.