src: fix sqlite connection leak in Web Storage

[JosephDoUrden]

Fixes #64640.

Every failed Web Storage initialisation leaks the SQLite connection. Root cause is in the issue thread: sqlite3_open() usually returns a handle even on failure, but Storage::Open() only adopts it into the RAII conn_unique_ptr on the last line of the function, so every early error return in between leaves the connection open. And since db_ never gets set on failure, the next localStorage operation retries the whole init and leaks another one. Measured before the fix: 50 failed inits on a corrupt file leaked exactly 50 fds; the later error paths (schema version check, empty state table) leaked 2 per attempt because WAL is established by then. After the fix all paths are flat.

The change adopts the handle immediately after sqlite3_open(), so RAII covers every return path. Two adjacent defects in the same function are included because a reviewer would trip over them anyway:

The sqlite3_prepare_v2() return value was overwritten before being checked by what looks like a copy-pasted second sqlite3_exec() of the init SQL, which also ran the whole script twice on every open. The stray exec is deleted and the prepare result is now checked. Schema output of a fresh db is byte-identical before and after (compared sqlite_master dumps), so running the script once is behaviourally inert.

The schema version read used CHECK(sqlite3_column_type(...) == SQLITE_INTEGER), which aborts the whole process if a pre-existing nodejs_webstorage_state table has that column typed differently (CREATE TABLE IF NOT EXISTS preserves whatever is there). A crafted or corrupted localStorage file should not be able to take down the process, so it now throws ERR_INVALID_STATE like every other failure in this function.

The test covers five failure paths (corrupt file, unopenable path, newer schema version, empty state table, non-integer schema version), each asserting the right error and a flat fd probe. The probe uses the lowest-available-fd trick (openSync on process.execPath) rather than listing /dev/fd, so it works anywhere with POSIX fd semantics; Windows is skipped because SQLite uses raw HANDLEs there and no fd probe can see the leak. The empty-state-table case doubles as a guard on the destruction order of the connection and statement holders, which the close-time CHECK_EQ(sqlite3_close, SQLITE_OK) depends on.

One thing deliberately left out: GetAll() also fails to check its sqlite3_prepare_v2() result, but the consequence there is an empty result rather than a leak or crash (sqlite3_step(nullptr) returns SQLITE_MISUSE and the loop never runs), and fixing it properly needs an error channel the current return type does not have. Separate PR.

AI disclosure: I used an AI coding agent for parts of the investigation and drafting. I verified the root cause against the SQLite and Node sources myself, and every number above comes from runs on my own machine.

[MikeMcC399]

Please make sure you have read and understood the following documents:

• guide for first-time contributors
• contribution and automation policies
• pull request guide
• AI use policy
• Code of Conduct

The guide for first-time contributors says that you should not open any new PRs until your first PR has been approved.

Your first PR in this repo, #65634, is awaiting review, so it has not reached the stage where it could be approved.

So long as you don't have a completed PR merged into the main branch, every PR change requires an explicit approval, as explained in Q: How do I trigger the CI runs? which from your point of view means additional delays, and from the collaborator / triager point of view means extra effort to monitor the PR and approve each run.

Waiting until your first PR has cleared also makes sure that you have experienced the complete process from start to finish.

Before you submit any further changes, you should test these locally. See https://github.com/nodejs/node/blob/main/doc/contributing/pull-requests.md#step-6-test.

[JosephDoUrden]

Thanks @MikeMcC399, understood. I'll hold any new PRs until #65634 clears. The format-cpp failure is fixed, ran make format-cpp locally and it produces no diff now