fix(agents): persist the catalog the availability probe already fetched

[kaizhou-lab]

Problem

The agent picker shows no modes / models / config options for agents that have never been opened in a conversation — including ones the availability check reports online.

The availability probe is not a shallow liveness check. It deliberately opens a real session/new so an auth-gated agent surfaces auth_required instead of silently looking healthy. That response carries exactly the catalog the picker needs — and the probe threw it away:

// custom_agent_probe.rs
let outcome = match protocol.new_session(...).await {
    Ok(_) => ProbeOutcome::Ok,   // ← modes / models / configOptions discarded

The four catalog columns were filled only when a live conversation opened a session and emit_snapshot_events broadcast to CatalogForwarder. So the data was fetched, proven, and dropped; the picker stayed empty until first use.

Measured on one dev machine: 27 of 39 ACP rows had config_options / available_modes / available_models NULL, several of them probing online. Verified the data is really there — a live probe of kilo returns a full configOptions array (model / effort / mode with all options) while its stored config_options is NULL.

Fix

Keep the response and persist it through the same channel a live session uses.

• catalog_partial_from_session_new (catalog_forwarder.rs) projects modes / models / config options into an AgentHandshake directly. It sits beside catalog_partial_from_event, and a parity test feeds one session state through both routes and asserts the blobs are byte-identical — both write the same agent_metadata columns, so which route ran last must not change what the picker reads.
  available_commands has no counterpart here: it arrives as a session/update notification, never in the session/new response.
• try_connect_custom_agent_with_catalog returns that partial alongside the verdict. The existing try_connect_custom_agent signature is unchanged, so the try-connect endpoint and test-on-save are untouched, and no SDK type leaks into aionui-api-types.
• The availability probe sends it via registry.catalog_sender(). The registry parameter was already threaded into run_probe and unused (_registry).

Safety

Additive by construction:

Case	Behaviour
Agent advertises nothing	Projection returns None — nothing is sent
A real session already filled the catalog	apply_handshake skips None fields, so a probe never blanks it
Agent not logged in	LIVE-verified: stakpak / poolside answer session/new with -32000 Authentication required and no payload. The Auth branch yields None, so unauthenticated agents correctly stay empty until logged in and re-probed

No schema change

The columns and the write path already exist; this only stops discarding data. Values land at runtime on the next probe — better than a seed migration, since a probe reads the catalog of the version actually installed on the user's machine rather than pinning one capture.

Tests

• probe_projection_matches_event_projection (new) — probe route vs event route produce identical blobs for the same session state; also pins that an empty/absent catalog produces no write
• catalog_partial_covers_session_fields — unchanged, still pins the event route

84 targeted tests green (catalog / availability / probe / registry). cargo clippy and cargo fmt clean. Full cargo nextest intentionally not run.