feat(cli): add cargo sqlx prepare --per-crate

[macalinao]

Adds cargo sqlx prepare --per-crate, which generates a .sqlx directory next to each workspace crate's Cargo.toml instead of a single one at the workspace root. Editing one crate's queries then only touches that crate's query data.

cargo sqlx prepare --per-crate
cargo sqlx prepare --check --per-crate

How

The read side already supported this. expand_input looks for each query file in $CARGO_MANIFEST_DIR/.sqlx before falling back to the workspace root, and it does so per file, so a crate can use either layout (or both). Only the write side was workspace-wide, because prepare sets one SQLX_OFFLINE_DIR for a single cargo check covering the whole workspace.

The obvious alternative — loop over members running cargo check -p <member> with SQLX_OFFLINE_DIR=<member>/.sqlx — is wrong: when crate A is checked, its workspace-member dependencies compile for the first time in that same invocation, so their query data lands in A/.sqlx. Avoiding that needs a warm-up build plus reverse-topological ordering, and N cargo check invocations.

Instead the macros route their own save directory. sqlx-cli sets SQLX_OFFLINE_PER_CRATE, and each expansion saves to $SQLX_OFFLINE_DIR/$CARGO_PKG_NAME. That is correct regardless of build order and still needs only one cargo check.

The staging tree lives under target/, and the CLI moves each crate's files into its .sqlx afterwards. --check reuses the same layout to compare against the checked-in data without touching it, and attributes failures to the crate they came from:

error: prepare check failed: .sqlx in crate-a is missing one or more queries; you should re-run sqlx prepare

Notes

• A query used by two crates is stored once per crate. That is inherent to the layout.
• Only query-*.json files are written or deleted, so anything else a user keeps in .sqlx survives. A .sqlx directory emptied of query data is removed if nothing else is in it.
• --per-crate implies --workspace and rejects --all, since crates outside the workspace have no crate directory to write to.
• After a successful run, query data left at the workspace root by a previous --workspace run is reported as no longer generated or checked, since the macros still fall back to it.
• save_in now creates its target directory, so a dependency outside the workspace that expands a query mid-prepare doesn't fail the build.

Testing

• 14 unit tests in sqlx-cli/src/prepare.rs covering the crate/directory mapping, staging reset, install (stale replacement, preserving non-query files, creating and removing directories), the check comparison, and the stale-workspace-data detection.
• 4 unit tests in sqlx-macros-core/src/query/metadata.rs covering save-directory resolution.
• 2 integration tests in sqlx-cli/tests/prepare.rs for the flag validation.
• Manually verified against a two-crate SQLite workspace: --per-crate writes each crate's queries to its own .sqlx with a shared query duplicated in both; SQLX_OFFLINE=true cargo check builds against them; --check passes when current and fails naming the right crate after a query changes; the stale-workspace warning fires; and --workspace still produces a single root .sqlx and passes --check.

https://claude.ai/code/session_011CmjiD9jHNqmdTptXkMu4j