Rebuild tool catalogs by upsert-then-prune so reads never see an emptied catalog

[ian-pascoe]

Summary

Tool-catalog rebuilds used to delete every tool and definition row for a connection and then re-insert them. D1 has no interactive transactions (interactiveTransactions: false), so each statement commits on its own:

• a search during a rebuild found zero tools for that connection;
• a rebuild that was cut off or overlapped another session's rebuild left a partial catalog. In production that was a stamped cloudflare_com catalog with 42 of 3600 tools.

The rebuild is now upsert-then-prune:

1. upsert definitions, then tools, on the tables' existing unique keys (no migration; created_at now means "first seen");
2. prune by name diff, reading names only, with deletes batched under D1's parameter limit;
3. stamp last, so an interrupted rebuild leaves every row intact and stays stale.

Definitions are written before the tools that $ref them and pruned after the tools that stop referencing them. Rows go in calls bounded by serialized size: fumadb runs a multi-statement D1 upsert as one native batch, and Cloudflare's catalog (37 MB) exceeded D1's 32 MiB RPC limit (#2123). The empty-catalog clears and removal cascades are unchanged.

Related: #2122, #2123.

Verification

New catalog-persist.test.ts runs rebuilds on the D1 adapter options, with an adapter proxy that pauses or fails writes and enforces a scaled-down batch payload cap:

• a read paused right after the rebuild's first tool write still sees the full previous catalog; afterwards, added, removed and updated tools and definitions are exact and the connection is stamped;
• rebuilding an unchanged catalog issues zero deletes;
• a rebuild whose inserts fail keeps every row and leaves the connection unstamped;
• a catalog larger than one batch payload lands in full.

All four fail against the previous implementation: the mid-rebuild read got [], the unchanged rebuild deleted tool and definition, the failed rebuild emptied the catalog, and the oversized catalog kept 1 of 400 tools.

Production (self-hosted Cloudflare deployment, cloudflare_com stale-marked, then 20 searches during the rebuild): the first search saw the previous 42 tools (never 0), and every later search saw 3600. D1 afterwards: stamped, 3600 distinct tools, 6900 definitions, with no exceeded* outcomes and no sync failures in wrangler tail. In 40 back-to-back searches across GitHub, Stripe and Cloudflare, 1 returned session_reset, down from ~50% before.

• oxfmt --check .
• oxlint -c .oxlintrc.jsonc . --deny-warnings
• typecheck (tsgo --noEmit) in core/sdk
• tests: core/sdk 947; also on the combined stack: plugins/openapi 338, plugins/mcp 311, plugins/graphql 118

Checklist

• Added a changeset.
• Added or updated tests for the new behaviour.
• No secrets, credentials, or private data in the diff.