fix: retry provider discovery periodically instead of once at startup

[AditiSah05]

Problem

Provider discovery (discover_all()) ran exactly once, at engine startup (engine.rs, in the constructor). If a provider — e.g. Ollama — wasn't reachable at that exact moment, the engine marked it unavailable and never checked again.

Consequence: start Ollama after the engine (slow boot, forgot to start it, container startup-order race), and /v1/capabilities stays empty forever. The only way out is a manual POST /v1/discovery/refresh call or a full engine restart — neither is discoverable to a normal user, and both contradict the project's own "zero-config, auto-detects, auto-recovers" pitch in the README.

The engine already runs two other background loops that repeat forever — spawn_idle_eviction() (unloads idle models) and spawn_artifact_sweep() (cleans old files). Discovery had no equivalent.

Fix

• mofa-engine-core/src/engine.rs

  • New spawn_discovery_ticker() — same pattern as the two existing tickers: a tokio::spawn loop on a tokio::time::interval, calling refresh_resources() (discovery + health) every tick.
  • Called at the end of the constructor, alongside the other two ticker spawns.
  • New fields discovery_task (task handle) and discovery_interval; both cleaned up in Drop (task aborted on engine shutdown, matching existing tasks).
  • New test discovery_ticker_picks_up_a_provider_that_was_down_at_startup: a fake provider (FlakyDiscoveryProvider) fails discovery until flipped "available" mid-test; the test asserts the model appears in capabilities() on its own, with no manual refresh_resources() call after the initial startup check.
  • Mechanical: build_engine_with_pricing test helper gained a discovery_secs parameter (3 call sites updated); a few explicit EngineConfig { ... } test literals needed the new field.

• mofa-engine-core/src/config.rs

  • New DiscoveryConfig struct: refresh_secs: u64, default 30, 0 disables the background refresh (same "0 = off" convention as idle_timeout_secs and artifacts.retention_secs).
  • Wired into EngineConfig as a new discovery field.

• mofa-engine-sdk/src/client.rs, mofa-engine-sdk/src/server.rs

  • No behavior change — test-only EngineConfig { ... } literals updated to include the new field so the crates still compile.

• config.example.toml

  • New [discovery] section documenting refresh_secs.

Testing

• cargo test -p mofa-engine-core — 179 passed. (9 pre-existing failures are unrelated to this change: local TTS/ASR/image-gen fixture tests spawn sh, which isn't on PATH on this Windows dev machine.)
• cargo test -p mofa-engine-sdk — 17 passed, including the existing auth_gate_rejects_missing_token_but_allows_public_routes test (unaffected).
• cargo build --workspace — clean.

Live-verified end to end, not just unit-tested:

1. Ran the built engine with Ollama genuinely down. Log showed the discovery warning repeating every refresh_secs (5s in this run), not a single one-shot warning as before.
2. Brought a fake Ollama HTTP responder online on port 11434 mid-run, with the engine still running (no restart).
3. Within one tick interval, the log showed discovered 1 models from 'ollama' and /v1/capabilities returned the model — confirmed in the dashboard UI too.

Non-goals

• No UI/dashboard code changed — the dashboard already polls /v1/capabilities; this PR just makes the data behind that endpoint stay fresh.
• No change to the existing manual POST /v1/discovery/refresh endpoint — still works, still useful for an immediate on-demand check.