[luthermonson]

Problem

The workspace entry for litewire-backend inherits that crate's default features, and its default is rusqlite. Every sibling crate (litewire-session, -mysql, -postgres, -tds, -hrana) depends on it through the workspace entry, so they all switch rusqlite back on. A downstream build that asks for

litewire = { ..., default-features = false, features = ["mysql", "postgres", "tds", "hrana", "turso"] }

still compiles rusqlite with bundled, which links a whole SQLite 3.46.0 and exports its sqlite3_* symbols.

ePHPm hit this. Its PHP pdo_sqlite/sqlite3 extensions link the PHP SDK's libsqlite3 by symbol name, and rusqlite's copy won the link, so PHP ran on a different SQLite than the one it was built against:

• ENABLE_MEMORY_MANAGEMENT: SQLite then uses one shared page cache with one process-wide mutex for every connection. With PHP running requests on 32 threads, a read-only FTS5 search endpoint did about 1,050 req/s, against about 9,000 req/s with the SDK's SQLite (numbers in fix(php): pdo_sqlite serializes under threaded PHP; run it on the SDK's SQLite ephpm#563, which bumps to this rev).
• DEFAULT_FOREIGN_KEYS=1: PRAGMA foreign_keys defaulted to on in PHP. Stock SQLite and stock PHP default it to off.
• No ENABLE_MATH_FUNCTIONS.

Change

litewire-backend = { path = "crates/litewire-backend", default-features = false } in [workspace.dependencies], plus a comment explaining why the flag matters.

The rusqlite backend is now opt-in only:

• litewire's backend-rusqlite feature, which is still in litewire's default, so the standalone crate behaves the same.
• An explicit features = ["rusqlite"]. The dev-dependencies already use this (litewire-session, litewire-mysql, litewire-turso), so the tests still get the backend.

Verification

• cargo tree -p litewire --no-default-features --features mysql,postgres,tds,hrana,turso -i rusqlite now reports that the package is not in the graph. Before this change it was there through every sibling crate. cargo tree -p litewire -i rusqlite with default features still shows it, through litewire.
• cargo check passes for each member crate on its own (-p litewire-backend/-session/-mysql/-postgres/-tds/-hrana/-turso), so no crate depended on the default without saying so.
• cargo clippy --workspace --all-targets -- -D warnings: clean. Also clean with ePHPm's feature set.
• cargo test --workspace --no-fail-fast: all pass except mysql_multi_tenant::single_tenant_attach_keeps_working_over_the_wire. That test fails the same way on main on Windows: the ATTACH path loses its backslashes and the test leaves UserslutherAppDataLocalTemp...other.db files in crates/litewire/. It is unrelated to this change, which I checked by running it with the previous Cargo.toml.
• In ePHPm with this rev pinned, libsqlite3-sys and rusqlite drop out of Cargo.lock, and PHP reports the SDK's SQLite 3.45.2 with foreign_keys = 0.

[ephpm]

ePHPm Preview — removed

Preview deployment has been torn down.