Rust-Embed Rebuild Invalidation

Problem

rfdb-server embeds the GUI bundle via rust-embed:

#[derive(RustEmbed)]
#[folder = "$GRAFEMA_UI_DIST"]
pub struct UiAssets;

You rebuild the GUI:

pnpm --filter @grafema/gui build       # produces dist/assets/HexAtlas-NEW.js
GRAFEMA_UI_DIST=$(pwd)/packages/gui/dist cargo build --release -p rfdb

Cargo says "Finished in 6 seconds" — way too fast for a full re-embed. You start the server, hit the URL, and curl http://localhost:51833/ui/default | grep HexAtlas returns the OLD hash. Re-running the build doesn't help.

Why this happens

rust-embed is a proc_macro evaluated at compile time. Cargo decides whether to re-run a proc-macro by hashing the source file that contains the #[derive(RustEmbed)]. Changes to environment variables (like GRAFEMA_UI_DIST) or to the contents of the asset folder do not invalidate the proc-macro. Cargo treats it as cache-hit, the previously expanded include-bytes! tree stays embedded, and your new asset never makes it into the binary.

Fix — touch the host file

touch packages/rfdb-server/src/static_ui.rs
GRAFEMA_UI_DIST=$(pwd)/packages/gui/dist cargo build --release -p rfdb

touch updates the file's mtime, which forces cargo to re-run the proc-macro, which re-reads the env var, which re-embeds the new assets. Build will take longer (full rust-embed re-expansion).

Verify:

strings target/release/rfdb-server | grep -E "HexAtlas-[A-Za-z0-9]+\.js" | head -3
# Should print the NEW hash (matches dist/assets/HexAtlas-NEW.js)

Why not cargo clean?

Works, but overkill — clears the entire target/, ~5 minute full rebuild. Touching one file is ~30 seconds and surgical.

Detection — how to know you've been bitten

If after cargo build --release:

1. ls -la target/release/<binary> — is the mtime older than your last pnpm build?
2. strings target/release/<binary> | grep <known-new-hash> — empty? You hit this bug.

Always verify embed freshness before debugging "the server returns old data". It's the second-most-common false-trail (after stale process on the port — see zombie-process-port-hunt).

Project script

For the Grafema build pipeline, the canonical recipe (from scripts/build-gui-for-rfdb.sh) should be:

pnpm --filter @grafema/gui build
touch packages/rfdb-server/src/static_ui.rs    # ← invalidate macro
GRAFEMA_UI_DIST=$(pwd)/packages/gui/dist \
  cargo build --release -p rfdb

Without the touch, the script's "rebuild rfdb-server with new bundle baked in" promise is silently broken on incremental builds.

Generalization

This pattern applies to any cargo proc-macro that reads files or env vars at expansion time:

• include_bytes! / include_str! — same problem if the included path is outside cargo's tracked tree
• serde_json::from_str(include_str!(...)) — same
• Custom build.rs that reads from $ENV and writes a generated source — usually OK if cargo:rerun-if-env-changed=ENV is declared, but easy to forget

If you see "fast cargo rebuild + stale embedded data" for any of these — the workaround (touch the host file) generalizes.