[spacedevin]

Motivation

The crates, rom and verify jobs run in ghcr.io/schlopai/chuggie-toolchain, which ships rustc, libmgba and agb-gbafix but no warm build. Every job (92 per full run) started with an empty registry and an empty target dir: crates.io index fetch, a clone of the agb fork, and a from-scratch compile of core/alloc (-Z build-std), agb and everything under it, before the example's own code was touched. Latest main run: the verify grid-demo "Run the verifier" step (build + 4000-frame emulation) is 1m44s of a ~2m50s job.

Change

crates

actions/cache@v6 over the registry (/usr/local/cargo/registry/{index,cache}, /usr/local/cargo/git/db — the image sets CARGO_HOME=/usr/local/cargo, not ~/.cargo) and crates/*/target (these crates are excluded from the root workspace, so each has its own target dir, as scripts/clean.sh documents).

rom and verify: one shared target dir

tish build normally writes each example's build under its own .tish/gba/<pkg>/target, a path no cache key could share across the 90 verify jobs, and CARGO_TARGET_DIR is deliberately scrubbed by tish (it breaks the ELF lookup; scripts/rom.sh documents the trap). tish has an explicit opt-in for exactly this case — TISH_GBA_TARGET_DIR (crates/tish_native/src/build.rs in @tishlang/tish 3.10.3: "share ONE cargo target dir across every example's ROM build … measured at ~45 GB across the fleet") — which points cargo and the ELF lookup at one directory. Both jobs set it to ${{ github.workspace }}/target/gba.

The dependency layer has the same fingerprint whichever example compiled it (same absolute paths for the path deps, same git rev for agb, same sysroot for build-std), so a target dir warmed by grid-demo is warm for every example.

• rom is the single writer (actions/cache@v6): it always runs, there is exactly one of it, and its post-step save only happens on success.
• verify jobs restore only (actions/cache/restore@v6) with the same key and paths. 90 jobs saving the same key would be 89 "cache already exists" warnings and 89 archives built for nothing.

Keys

crates-<sha256 rust-toolchain.toml>-<sha256 crates/**/Cargo.lock, .cargo/config.toml, package.json>
gba-target-<sha256 rust-toolchain.toml>-<sha256 .cargo/config.toml, package.json, crates/**/Cargo.toml>

with restore-keys on the <prefix>-<toolchain hash>- part. The nightly pin leads so a restore never crosses a rustc change (cargo would ignore those artifacts, and they would only bloat the next save). .cargo/config.toml carries the agb rev, package.json the exact tish version; package-lock.json is deliberately not in the key, since deps bumps like the deckfile chores would otherwise miss the cache for a change that does not touch the Rust graph.

What this does and does not cache

Cargo fingerprints path dependencies by mtime, and a fresh checkout / npm ci is always newer than the cached fingerprints, so crates/tish-agb, tish-gba-scenepack, tish-agb-sio, tish-gba-game-engine, node_modules/@tishlang/tish/crates/tish_runtime_gba and the game crate itself still compile in every job. What stays warm is everything beneath them: the registry index, the agb fork clone, core/alloc/compiler_builtins from build-std (rust-src in the image has fixed mtimes), agb and its crates.io dependencies, and scenepack's proc-macro dependencies (syn/quote/proc-macro2, host builds). That is the bulk of a cold build, but not all of it.

Caveats, stated plainly

• Billing is per job rounded up to the minute. A verify job that goes from 2m50s to 2m10s still bills 3 minutes. The real saving depends on where each job's wall time lands relative to the minute boundary; expect the first few runs to tell. Restore cost is one download + untar per job (the archive is probably several hundred MB), so the net is compile time saved minus that.
• The toolchain image has no zstd, so actions/cache will fall back to gzip in these jobs (save and restore are both in the container, so they agree). Adding zstd to the image's apt list (schlopai/chuggie-toolchain, not this repo) would make both ends faster.
• Nothing in the repo reads .tish/gba/<pkg>/target after this: the verifiers that reach into .tish/gba/ read src/main.rs and Cargo.toml (checked: grep -rn 'target/' examples/*/verify.sh scripts/*.sh scripts/*.py); check_agb_fork.sh uses cargo metadata, which does not involve the target dir. The two verify.sh that unset CARGO_TARGET_DIR do not touch TISH_GBA_TARGET_DIR.
• Swatinem/rust-cache was not used: it assumes ~/.cargo and a workspace target/, neither of which is the layout here.

Both files parse (actionlint, YAML load); the one SC2155 note actionlint reports is pre-existing on main.