fix(oracle): a failed connect closes the pool it created (#1102)

[sloemo01]

Resolves #1102.

What this is

OracleProvider.connect() built its pool, then tested it with a borrow. When that borrow failed, the catch rethrew without closing the pool and without clearing this.pool. factory.getOrCreateProvider() never caches a provider whose connect() threw, so nothing held a reference that could close it, and node-oracledb's background creator keeps reaching for poolMin connections with no delay between attempts — a spinning core and a connection flood that outlives the process's usefulness.

The catch now does what PostgreSQL and SQL Server already do:

const failedPool = this.pool;
this.pool = null;
await failedPool?.close(0).catch(() => {});

close(0) is the oracledb equivalent of end(). Clearing this.pool matters as much as closing: left set, the if (this.pool) guard at the top of connect() returns on a retry without dialling and without an error, so a caller that retries reads a silent success. The ticket lists this as a second, smaller defect from the same line, and one test arm covers it directly.

A close() that itself rejects is swallowed, so a cleanup failure never becomes the reason a connect was refused.

Tests, written first

Four arms in tests/integration/db/oracle-provider.test.ts, using the file's existing mockCreatePoolFn / mockPoolCloseFn doubles:

arm	before the fix
a getConnection failure closes the pool it just created	close called 0 times
the NJS-138 config path closes the pool too	close called 0 times
a connect retried after a closed failure creates a new pool	second connect() returned early
a close that itself rejects does not replace the connect error	close never attempted

All four failed on current main (209 pass, 4 fail), all four pass with the fix (213 pass). The arms count close calls rather than asserting on this.pool alone, because a pool can be unreferenced and still running — which is the defect, and the reason a null check would have gone green on the broken code.

Two of the arms pin things the acceptance criteria ask for explicitly: the NJS-138 path (mapped instanceof DatabaseConfigError throws before the fix's code would have run in a naive ordering) and the error-identity rule that a rejected close must not replace the original connect error.

Docs

docs/providers/oracle.md gains §3.6, recording why a failed connect closes its own pool and the measured cost of not doing it, per the provider triad rule in CLAUDE.md.

Verification

• bun tests/run-tests.ts tests/integration/db/oracle-provider.test.ts — 213 pass, and 100% line coverage on src/lib/db/providers/sql/oracle.ts (1012/1012, zero uncovered lines) from the lcov this PR produces.
• bun run format clean, bun run lint 0 errors, bun run typecheck clean.

One thing worth stating plainly rather than burying: the repo's full suite does not go green in this environment, and not because of this change. On a clean origin/main with this branch absent, these files already fail: sqlite-provider (83), sqlite-driver (27), launcher-utils (6), object-edit-declarations (4), result-pagination-capability (4), object-source-declarations (3), object-column-declarations (3), agent-investigation-e2e (2), test-runner-cli (2), monaco-language-ids (1), agent-statement-boundary (1), and mongodb-provider exits without writing a report. The identical list appears with my change applied. Since those abort bun run test:coverage before the lcov merge, I verified the coverage gate for the file this PR touches instead of claiming the global gate passed.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[cevheri]

great, thanks
just one docs error(reference) : docs/providers/oracle.md:1572 #36-privilege-resilient-monitoring

[sloemo01]

Fixed in 3bb16ca.

You are right, and it was a miss on my side: inserting the new section made the monitoring one §3.7, and the reference in §8 kept the old target. #36-privilege-resilient-monitoring no longer matches any heading, so the link went nowhere.

One line, at docs/providers/oracle.md:1587: §3.6 -> §3.7, #36-privilege-resilient-monitoring -> #37-privilege-resilient-monitoring.

Checked the rest of the file while I was in there: every other internal anchor resolves against a real heading, and there are no remaining §3.6 text references. The section itself is unchanged.