Bind a lone buffer positionally instead of as named parameters

[ivanjeremic]

The bug

Binding a single Buffer to a statement panics out of the runtime:

const db = new Database(":memory:");
db.exec("CREATE TABLE kv (key BLOB PRIMARY KEY, value BLOB NOT NULL)");
const key = Buffer.from([1, 2, 3]);

db.prepare("INSERT INTO kv (key, value) VALUES (?, ?)").run(key, Buffer.from([9])); // fine
db.prepare("SELECT value FROM kv WHERE key = ?").get(key);                          // panics

thread '<unnamed>' panicked at src/lib.rs:1354:62:
called `Option::unwrap()` on a `None` value

On 0.6.0-pre.41 it aborts the process outright (fatal runtime error: failed to initiate panic).

Why

map_params decides positional vs named by JS type, and a Buffer is an object:

ValueType::Object => {
    let object = params.coerce_to_object()?;
    if object.is_array()? { map_params_array(object) } else { map_params_object(stmt, object) }
}

So a lone buffer goes to map_params_object, which asks for each parameter's name. For an anonymous ? that is None, and the .unwrap() panics. Two arguments take the positional path, which is why the insert above survives and the select does not — and why the existing blob tests pass: they bind inside an array (insertStmt.run([array])), never alone.

The fix

• Buffers and typed arrays take the positional path, as they already do inside an array. map_value has the is_buffer / is_typedarray checks; this reuses them one level up.
• map_params_object reports an anonymous-parameter statement instead of unwrapping None, so genuine misuse — a named-parameter object against ? placeholders — throws a JS error rather than killing the process.

Tests

Added to integration-tests/tests/sync.test.js:

• a lone Buffer, a Buffer inside an array, and a Uint8Array all bind and read back the same blob;
• a named-parameter object against anonymous ? throws rather than aborting.

I could not build or run the suite locally — no Rust toolchain on the machine I wrote this on — so the change is source-reviewed only and needs your CI to confirm it. Reproduced the bug itself on both 0.5.29 and 0.6.0-pre.41.

Related

• Statement.run() with object parameter panics #114 (Statement.run() with object parameter panics) is this same unwrap, reached the other way, and the second change turns it into an error.
• Panic (unwrap on None) reading sqlite-vec vec0 KNN results — SELECT ... WHERE embedding MATCH ? #229 and Storing blobs causes a panic #60 look like the same root cause — both bind a single blob-ish value — though I have not verified either repro.

Found while using libSQL as a key-value engine, where every key is a Buffer; the workaround until now was encoding keys as hex text.

[ivanjeremic]

Update: I've now built this locally (Rust 1.98.1, npm run build:debug) and verified it against the package entry point, so the "source-reviewed only" caveat in the description no longer applies.

Before the patch, the second line below aborted the process. After it:

insert with a Buffer: ok
select by a lone Buffer: { value: <Buffer 09> }
select by a lone Uint8Array: { value: <Buffer 09> }
range scan by Buffers: 1 row(s)
named object against a ? statement: SQLite3 cannot bind named parameters to a statement using anonymous '?' parameters; pass them positionally instead

So both halves hold: buffers and typed arrays bind positionally, and a named-parameter object offered to a statement with anonymous ? throws a JS error rather than panicking out of the runtime.

I have not run the full integration-tests suite locally — it expects a provider setup I don't have — so the two tests added here still need your CI.