fix(sql-editor): repair Server Beam gates, and make source/target visible

[huyplb]

Review of #158. The feature's design is sound — but it landed with the gates red, and the source/target mapping is invisible in exactly the operation where getting it wrong is worst.

main was shipping broken

Verified by stashing my changes and running against clean main:

Gate	State on main
cd apps/web && npx tsc --noEmit	2 errors (TS2456, TS2577)
npx vitest run	2 failing tests

• SqlBinding = ReturnType<typeof makeSqlBinding> on a function annotated : SqlBinding is circular. Dropped the annotation, let it infer; the alias still serves consumers.
• server-beam.test.ts expected /at most 2/ but the message read "cant handle more than 2". Aligned the message (and its missing apostrophe).
• codeCellExec.test.ts still asserted the old contract. feat: Server Beam — sql.on across two Destinations #158 deliberately widened normalizeCodeCellReturn so return 1, return [1,2,3] and a bare object all become grids — a good change. I updated the assertions to the new behaviour rather than narrowing the feature back.

Alias lookup accepted inherited keys

code-cell-thread.ts guarded with !beamDialects[key]:

toString     passes guard: true   typeof: function   ← wrong
constructor  passes guard: true   typeof: function   ← wrong
__proto__    passes guard: true   typeof: object     ← wrong
nope         passes guard: false                     ← correct

sql.on('toString') skipped "Unknown Server Beam alias" and used the inherited function as the dialect, dying downstream as dialect.toLowerCase is not a function. Now Object.hasOwn.

Not exploitable — the parent routes through a Map and fails closed — but a confusing dead end. This is the third appearance of this class here (after setBinding in the import parser); a shared hasOwn helper or a lint rule would be cheaper than a fourth fix.

Which server is target was invisible

This is the one that can lose data. Aliases come from list order, not click order, and sql.on('target') is what writes:

• A third checked Destination was silently dropped by slice(0, 2) — so parseBeamEndpoints' own "at most 2" error could never fire from the UI. Now an explicit error telling the user to uncheck the extras.
• Every beam run now prints the resolved mapping — Server Beam → source = A, target = B — before results.

The server side needed nothing: aliases resolve per-user via resolveRef, each shares the same permission policy from #154, and dialects are per-alias. Credit where due — a second connection path is where a gate usually gets forgotten, and it wasn't.

Samples — one per editor case

Both executed against real SQLite before committing:

• General (one server, no alias) — plain sql\…``, no beam. Verified: returns the bound row.
• Migration (source → target) — read, reshape, chunked write, read back. Verified across two separate database files: three rows landed in target with o'brien@example.com intact, and sqlite_master confirmed the source was untouched.

Writing that sample caught a bug in the sample itself — domain was split from the pre-lowercased email, producing Example.COM next to o'brien@example.com. Normalize once, then derive.

801 tests pass, tsc and ESLint clean.

Note on the earlier hang

My first attempt at the migration sample hung for two minutes. That was my harness, not the product — I had run it before seeding the source database. Isolated it by testing the beam machinery with stubs, the runners standalone, and one sql.on through the worker, all of which passed; the full sample then ran fine once both files existed. Worth knowing the shape though: because the cell clock pauses during a bridged query, a runner that never settles would hang without ever hitting the timeout.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes affect Server Beam migration writes and alias routing; mistakes could target the wrong database, though the PR adds explicit mapping warnings and stricter destination limits.

Overview
Repairs Server Beam behavior and visibility in the SQL editor: alias resolution, destination selection, and documentation samples.

Server Beam alias lookup in the code-cell worker now uses Object.hasOwn on the alias→dialect map so inherited keys like toString are rejected with a clear unknown-alias error instead of failing later on dialect handling. Tests document that trap.

Execute path no longer silently keeps only the first two checked Destinations when more are selected; runs warn and stop when too many are checked, and each beam run prints Server Beam → source = …, target = … so list order (not click order) is explicit before writes on target. makeSqlBinding drops an explicit return type that caused a circular TypeScript error.

Tests for code-cell return normalization now expect scalars, scalar arrays, and bare objects as grids (only null/undefined rejected). Sample bookmarks add one-server general Node SQL and a source→target migration example. Server Beam parse error text is aligned with the “at most 2” cap.

^{Reviewed by Cursor Bugbot for commit 286a485. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_2536d2ed-e5f0-4d70-9e8f-3a63d7f85fee)