CI takes 30-40min due to macOS Nix cache and redundant sol-prelude

[thedavidmeister]

Problem

CI wall-clock time is 28-40 minutes, dominated by two bottlenecks that aren't actual test/build work.

Step-level breakdown (from run 24310718017)

macOS jobs (the overall bottleneck)

Step	rs-artifacts	rs-test
Nix cache restore	1069s (18m)	1178s (20m)
sol-prelude	174s (3m)	188s (3m)
actual work	365s (6m)	187s (3m)
checkout	118s (2m)	88s (1.5m)
total	1753s (29m)	1671s (28m)

Ubuntu jobs (for comparison)

Step	rs-artifacts	sol-test	rs-test
Nix cache restore	34s	39s	36s
sol-prelude	551s (9m)	554s (9m)	563s (9m)
actual work	403s (7m)	137s (2m)	197s (3m)
checkout	64s	88s	79s
total	1071s (18m)	837s (14m)	895s (15m)

Key findings

1. macOS Nix cache restore: 18-20min vs ubuntu 34s (30-50x slower)

The Nix cache restore on macOS takes 18-20 minutes — that's 61-71% of the macOS job time. On ubuntu the same step takes 34-39 seconds. The actual test work on macOS is only 3-6 minutes.

2. sol-prelude (forge compile) runs identically in all 8 jobs

Every single job runs rainix-sol-prelude independently — 9 minutes each on ubuntu, 3 minutes on macOS. This is the same compilation producing the same artifacts 8 times. Could be factored into a shared dependency job whose artifacts are downloaded by downstream jobs.

3. Actual test/build work is fast

Job	Actual work
sol-test	2m 17s
sol-static	22s
rs-test (ubuntu)	3m 17s
rs-test (macos)	3m 7s
rs-artifacts (ubuntu)	6m 43s
rs-artifacts (macos)	6m 5s
rs-static	2m 1s
wasm-build	1m 27s

Potential improvements

1. Fix macOS Nix cache — investigate why cache restore is 30-50x slower than ubuntu. This alone would cut macOS jobs from ~30m to ~10m.
2. Share sol-prelude — run it once as a dependency job and pass artifacts to downstream jobs. Saves ~9min per ubuntu job (currently wasted on redundant compiles).
3. Question macOS necessity — if macOS jobs exist only for Rust cross-compilation, consider whether that can be done on ubuntu instead.