[sanjevirau]

Faster CI and release runs, and the two Auth tests that kept making a re-run necessary. Part of the 0.2.4 milestone.

What changes

• The engine no longer rebuilds on every cargo build in a checkout without a console build. crates/console-front/build.rs watched console/dist/index.html. Cargo re-runs a build script on every build while a watched path is missing, so console-front, suite-runtime and the CLI were recompiled each time: about 43 s on a CI runner, in most steps of the Differential harness and the client replays. Locally, a no-change build goes from 5.6 s to 0.15 s; adding, changing or removing a console file still re-embeds it.
• CI: the sign-in p99 test fails now and then on the Intel macOS runner #190, the sign-in p99 test measures up to three rounds and passes on the first within both bounds. It failed release 0.2.3's first attempt and chore(release): prepare 0.2.3 (engine 2294132, the Indexes page, three reported bugs fixed, a container image) #184's and chore(deps): bump notify from 8.2.1 to 8.2.2 in the rust-dependencies group #185's packages runs. The persistence test flushes before removing the state directory, so the writer thread can't add a file to it meanwhile ("Directory not empty").
• darwin-x64 is built on Apple silicon for x86_64-apple-darwin (checked with lipo), and its release-mode tests, npm and Bun installs and suite smoke run the x86_64 engine with an x64 Node under Rosetta 2. The Intel runner took 20–47 min for this job, twice the next slowest platform, and set the length of every packages and release run.
• The draft GitHub release gets this version's changelog section (packaging/release-notes.mjs), not the whole changelog. On 0.2.3 it gives exactly the notes that were published.

Checked locally: the Auth tests in release (82 passed, the two changed tests 5 more times each), node --test packaging/*.test.mjs (195 passed), node packaging/check-release.mjs, the independence check, and the build script's rebuild behaviour in a fresh checkout, after a console build appears, after a nested asset changes and after it is removed.

Before/after run timings follow in a comment once this PR's runs finish.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GojiTpvSY8YywZFwh2ZJg2

[sanjevirau]

Before and after. Wall time is from the run's start to its last job. "Before" is the last five runs of each workflow on main or a code PR before this one.

	Before	After
CI	25.1–31.0 min (median 30.6)	13.9 and 13.4 min (two runs)
— the build-script fix alone		20.1 min
Differential harness	24.6–30.5 min	9.8 and 8.3 min, with the oracle replays (12.0 and 11.7) beside it
Client replays and real SDK checks	15.3–18.0 min	8.2 and 8.9 min
CI runner time	155.7–167.5 min	143.9 and 144.7 min
Installable packages	35.4–38.5 min (median 37.0)	17.6 min
— darwin-x64 job	30.1–37.9 min (46.8 in release 0.2.3)	17.2 min; its release build 8 min 40 s, against 18–25 on the Intel runner
Packages runner time	81.5–93.2 min	68.7 min

A release runs CI and the packages together, as a pull request does, so its build phase before the approval should go from 34.8 min (0.2.2) and 35.4 (0.2.1) to about 18. For 0.2.3 that phase took 76.4 min, because #190 failed the first attempt. That figure is an estimate from this PR's runs; a release can't be dry-run.

The darwin-x64 job on this PR ran on the macos-15-arm64 image with Node darwin-x64 v24.20.0. lipo reported x86_64, the release-mode tests passed on the x86_64 binaries (82 Auth, 1 suite control), and the npm, Bun and suite smokes installed and ran @firenook/cli-darwin-x64.

Two failures here were not caused by this change. linux-x64's old-glibc step hit a Docker Hub 502 Bad Gateway pulling node:24.20.0-alpine; it passed on a re-run. CI's second run failed functions-invocations.spec.ts:38 once ("Show" detached while the feed re-rendered); the same test failed on 2026-10-27 in run 40599798543.