Evaluation failure detail lives in scenario_results JSONB; no error column, no failure reason on the API

[emooreatx]

For status='failed' rows there is no dedicated error field. The failure is stored as scenario_results = {"error": …, "first_error": …} with total/correct/errors NULL — the same column that holds per-scenario results for successful runs. Examples from 2026-02-16 (host inventory 2026-09-24):

• gemini-2.5-pro: "Catastrophic error rate: 277/300", first_error "Empty response (finishReason=MAX_TOKENS)"
• openrouter-llama-4-maverick: HTTP 400 "meta-llama/llama-4-maverick-instruct is not a valid model ID" (300/300)
• together-deepseek-v3.1: HTTP 404 "Unable to access model deepseek-ai/DeepSeek-V3.1-0324" (300/300)
• gemini-2.5-pro ×4 total: HTTP 429 quota exceeded
• claude-sonnet-4-5 / claude-opus-4-6 (EU): HTTP 400 "Your credit balance is too low to access the Anthropic API"

Every one of these is an operational failure (bad model ID, quota, billing) that nobody saw for seven months because the only place it's recorded is a JSONB blob on a row the scores endpoint filters out.

Asks

1. Add error_code / error_message / error_class (provider_config | quota | billing | model_output | timeout) columns; keep scenario_results for results.
2. Surface failures: /api/v1/admin/frontier-sweeps/{id} should list failed models with error_class, and the admin UI Frontier page should show them red instead of just "N completed".
3. Sweep summary notification (email/Discord — services/notifications.py already exists) when any model fails with provider_config or billing — those are the two classes that will fail identically next week.

Related: #34, #39.

[emooreatx]

Concrete instance, 2026-09-25: after migration 022 the bench crash-recovery reaper ran for the first time and marked the 15 running rows failed (US now completed 59 | failed 19). The rows have status='failed', completed_scenario_count=0, completed_at set — and scenario_results NULL. The reason ("Server restart: 0/300 scenarios completed before crash") exists only in the bench container's stdout. So even the good path for failures leaves no queryable trace of why. Same root cause as this issue: there is no error_* column for the writer to fill.