Derive `catalog` from `mcp_url` via the MCP head's /version (server half now shipped)

[cboettig]

Breaking out item 3 of #335 ("optionally make catalog itself optional, so a headless/MCP-only config is just {"mcp_url": "…"}"), because the server-side blocker is now gone.

What changed on the server

mcp-data-server#346 / PR #348 makes an MCP head advertise the client-reachable URL of the catalog it reads, as structured data on a public endpoint:

$ curl -s https://duckdb-mcp.carlboettiger.info/version
{"version":"main","git_sha":"…",
 "stac_catalog_url":"https://minio.carlboettiger.info/public-data/stac/catalog.json",
 "public_base_url":"https://duckdb-mcp.carlboettiger.info"}

Properties that matter for this repo:

• No MCP session, no auth token. /version is deliberately auth-exempt, so this works before mcp.connect() resolves and on token-gated heads.
• Browser-reachable. Verified access-control-allow-origin: * on both heads today (duckdb-mcp.nrp-nautilus.io and duckdb-mcp.carlboettiger.info), so no proxy is needed.
• Always public. stac_catalog_url is the client-facing URL even when the server itself reads an in-cluster address — that was the whole point of #346. Defaults to the server's own catalog URL, so heads on NRP prod already answer usefully.
• Only public URLs are exposed there — the in-cluster address stays inside authenticated MCP responses.

Proposed change here

app/main.js already resolves mcpUrl at step 1c (app/main.js:64), before catalog.load(appConfig) at step 2+3, and DatasetCatalog.load reads appConfig.catalog at app/dataset-catalog.js:52. So the hook is small:

1. When appConfig.catalog is absent, derive it: GET <mcpUrl without trailing /mcp>/version → stac_catalog_url. A headless/MCP-only config becomes just {"mcp_url": "…"}.
2. Optionally, when the configured catalog fetch fails after DatasetCatalog.load() still fatal after retry exhaustion (and fetches the root even with collections: []) — blocks a working MCP-only degraded mode #335's retry exhaustion, try the derived URL before degrading to an empty catalog — a head configured against a live mirror knows a reachable catalog even when the app's hardcoded one is down. That is exactly the case that cost the manual hunt in mcp-data-server#346.
3. Keep it non-fatal: if /version is unreachable or lacks the field (a head predating #346), fall through to DatasetCatalog.load() still fatal after retry exhaustion (and fetches the root even with collections: []) — blocks a working MCP-only degraded mode #335's degraded path. No new hard dependency on the boot critical path.

Ordering note: this composes with #335 but doesn't need it — 1 is independent, 2 slots into whatever #335 lands as the post-retry branch.

Scope caveat carried over from #335

A derived catalog is still not equivalent to the curated config: the app's collection list is what feeds list_datasets, get_schema, and the dataset paths in the assembled system prompt. Deriving the root URL doesn't change that — an MCP-only config still has collections: [] and still shifts agent behaviour onto the MCP STAC tools. Right for a smoke test or an outage, wrong for a benchmark meant to measure the production app.

Related: #335 (degrade instead of throwing; item 3 is this), #197 (source the catalog from MCP instead of walking STAC client-side), #287 (retry, shipped), mcp-data-server#346.