One discovery pass for every harness baseline

The previous goalpost, after the provider-aware harness_models schema landed, was still a static catalog. BUILTIN_MODEL_CATALOG and a single Agy live check were what dispatch could trust. Sub-task 1 stored rates, effort, and a (harness, model, provider) key. Nothing wrote those columns from a live probe.

What changed

run_model_catalog_discovery(conn) walks HARNESS_CAPABILITY_BASELINES instead of a fixed list of four CLIs. Each harness is recorded with provider='native'.

Claude, Agy, and Grok still use their model-list commands. The probe now keeps rate, context, and explicit tier fields when the listing payload actually contains them. Agy ids that end in -low, -medium, or -high set complexity_tier and max_effort_level. Any other id leaves both columns null. A curated row never receives a tier just because its id looks tiered.

Codex has no list command. The pass reads model from CODEX_HOME/config.toml and stores discovery_source='config_file'. If that file has no model, the harness falls back to BUILTIN_MODEL_CATALOG with discovery_source='curated_fallback'. The same fallback covers a listing harness whose probe returns nothing. When the probe names a model but does not quote a price, the catalog RateCard is scaled from per-1k to per-million tokens and stored with rate_source='curated_fallback'.

CONFIGURED_GATEWAY_PROVIDERS is empty. The gateway loop runs and queries nothing, so an OpenRouter row is not marked stale and no adapter is called. Previously active native rows that this pass did not see become status='stale'. They are not deleted.

synlynk models discover calls this function. resolve_tier_model, dispatch --model, and the 24h watch are unchanged. Those are the next two sub-tasks.

Goalpost

The catalog table can now be filled from whatever harnesses are registered, with rates and tiers only where a probe or the curated card actually supports them. The next step is the 24h refresh, then dispatch reading status='active' rows instead of the static catalog.