fix(v0.18.2.fork.1): v26 — heal string-encoded source configs before jsonb_set

Prod LXC 107 deploy of v26 (2026-05-07) failed with SQLSTATE 22023
"cannot set path in scalar" because 6 of 7 sources had jsonb_typeof = 'string'
instead of 'object'. Root cause is a pre-existing bug in sources.ts:211:

  await engine.executeRaw(
    `INSERT INTO sources (...) VALUES (..., $4::jsonb) ...`,
    [..., JSON.stringify(config)],
  );

postgres-js's unsafe() with $::jsonb cast double-encodes the JSON string —
the cast lands as a JSON STRING scalar, not the intended object. Migration-
inlined inserts (e.g. v17 'default' source) work correctly because they use
literal '{"key":"val"}'::jsonb at SQL level.

v26 was the first migration to hit jsonb_set on these legacy configs,
which is why this surfaced now (drill on D-LXC fixture missed it because
the fixture was empty + sources-add via CLI hit the bug but no further
jsonb_set ran on those rows).

Fix: prepend a Step 0 to v26 that unwraps any string-encoded config back
to its object form via (config #>> '{}')::jsonb. Idempotent on already-
object configs (filtered by jsonb_typeof). Byte-equivalent contents — the
JSON parse step is information-preserving.

Manual prod recovery (2026-05-07 14:05 UTC): unwrap UPDATE applied to LXC
107 BEFORE this commit, then v26 re-ran and applied cleanly. Post-state
verified: 203 gstack-brain pages → 155 stock-dashboard + 40 memory-dashboard
+ 8 default-ambiguous, gstack-brain source dropped, default-ambiguous +
gstack-meta sources created.

This fork commit codifies the fix so future Postgres deploys (other dev
boxes, fresh prod redeploys, the in-progress gbrain-mcp:v0.18.2-fork.1
image rebuild) self-heal automatically. Adds a regression test
(string-encoded config) in tests/migration-v26.test.ts. 14/14 tests pass.

Followup TODO: fix sources.ts:211 to either pass an object directly (let
postgres-js handle JSON serialisation) OR use postgres.json() helper.
Out of scope for this commit — the unwrap heals existing data; an
upstream fix prevents new corruption.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/core/migrate.ts

@@ -866,6 +866,20 @@ export const MIGRATIONS: Migration[] = [
     version: 26,
     name: 'source_taxonomy_rewrite',
     sql: `
+      -- Step 0: heal pre-existing data corruption from sources.ts:211.
+      -- gbrain CLI's runAdd writes config via $4::jsonb on a JSON.stringify()'d
+      -- value through postgres-js unsafe(), which double-encodes the payload —
+      -- the cast lands as a JSON STRING scalar instead of a JSON object.
+      -- Verified on prod LXC 107 (2026-05-07): 6 of 7 sources had jsonb_typeof
+      -- = 'string' (only the migration-inlined 'default' source was a true
+      -- object). jsonb_set() in subsequent steps fails on scalars with
+      -- "cannot set path in scalar" (SQLSTATE 22023). Unwrap is byte-equivalent
+      -- (parses the JSON string back to its underlying object form) and
+      -- idempotent on already-object configs (filtered by jsonb_typeof).
+      UPDATE sources
+        SET config = (config #>> '{}')::jsonb
+        WHERE jsonb_typeof(config) = 'string';
+
       -- Step 1: ensure tombstone source exists for slug-no-match fallbacks.
       INSERT INTO sources (id, name, config)
       VALUES (

test/migration-v26.test.ts

@@ -159,6 +159,37 @@ describe('v26 — page reclassification', () => {
   });
 });
 
+describe('v26 — string-encoded config heal (regression)', () => {
+  test('migration unwraps jsonb string scalar configs to objects before jsonb_set', async () => {
+    // Reproduces prod LXC 107 (2026-05-07) data corruption: gbrain CLI's
+    // sources.ts:211 INSERT via $::jsonb on JSON.stringify() output produces
+    // a JSON STRING scalar, not an object. jsonb_set on a scalar throws
+    // SQLSTATE 22023 'cannot set path in scalar'. v26 step 0 unwraps before
+    // the rest of the migration touches config.
+    await engine.executeRaw(
+      `INSERT INTO sources (id, name, config) VALUES
+        ('regression-string-cfg', 'regression-string-cfg',
+         '"{\\"federated\\":true,\\"slug_prefix_rules\\":[\\"regression/\\"]}"'::jsonb)
+       ON CONFLICT (id) DO UPDATE SET config = EXCLUDED.config`,
+    );
+    // Sanity: confirm we set up the bug condition.
+    const before = await engine.executeRaw<{ type: string }>(
+      `SELECT jsonb_typeof(config) AS type FROM sources WHERE id = 'regression-string-cfg'`,
+    );
+    expect(before[0].type).toBe('string');
+
+    // Re-run v26: step 0 should unwrap, then the remaining steps proceed cleanly.
+    await engine.runMigration(26, v26Sql);
+
+    const after = await engine.executeRaw<{ type: string; rules: string[] | null }>(
+      `SELECT jsonb_typeof(config) AS type, config->'slug_prefix_rules' AS rules FROM sources WHERE id = 'regression-string-cfg'`,
+    );
+    expect(after[0].type).toBe('object');
+    // Contents preserved byte-for-byte after unwrap.
+    expect(after[0].rules).toEqual(['regression/']);
+  });
+});
+
 describe('v26 — idempotency (CR-6)', () => {
   test('re-running migration is a no-op: source distribution unchanged', async () => {
     const before = await engine.executeRaw<{ source_id: string; n: bigint }>(