[BENCH-839] Drive the LLM roster from the models table

[coval-cale]

Follows #605, which made the LLM proxy, Coval agent, sync, and fetch provider-agnostic but left the provider list as a code constant. This replaces that constant with the LLM rows in benchmarks_v2.models, so the collected flag STT and TTS already use becomes the on/off switch for an LLM provider.

What changes

• sync-llm reads LLM rows from the DB, reconciles one Coval agent, run template, and schedule per row, and sets the schedule's enabled flag from collected. It skips the fetch for uncollected rows.
• LLM fetch builds its specs from collected rows only, and requires the dental test set up front.
• Proxy resolves /llm/{provider} against the roster on each request. No row or collected false is a 404. Row present but no client is a 503.
• Clients stay in code as a per-provider factory map in llm/benchmark.py. Startup logs llm_client_missing for any collected LLM row without a factory, which addresses the registry/client drift finding from Proxy and sync every LLM provider through one definition #605.
• Phonely's client gains a from_settings factory. Its HTTP behavior is unchanged.

What it enables

• Pausing Phonely is flipping collected to false on its row through the admin models endpoint. The next llm-sync (5:00 AM Pacific) disables the Coval schedule and stops fetching, and the proxy refuses its turns. Flipping back re-enables.
• A second LLM is a models row, a client class, one factory-map line, a coval_llm_<slug>_agent_id setting, and credentials.

Not changed

Database schema, config, the persona defaults from BENCH-825, and Phonely's live Coval agent. Deploys as a no-op for Phonely.

Tests pin the schedule disable/unchanged/enable transitions, the CLI skipping the fetch for a paused row, the proxy 404 for an uncollected row, and roster-built fetch specs.

Linear: BENCH-839