perf(ci): speed up PR CI wall clock and local dev builds

[wpfleger96]

Summary

Measured baseline: PR CI ~10 min wall clock, release.yml ~25 min wall clock.

CI — critical path & compute

• profile.ci opt-level 1 for workspace crates (Cargo.toml): deps keep opt-level = 3 via [profile.ci.package."*"] so the main-seeded dep cache stays byte-identical. The relay build should drop from ~4m38s → ~2m15–2m45s. opt-level 1 keeps debug-assertions off (release inherit) so the e2e timing stability from #307 is preserved.

• mold linker (rui314/setup-mold@v1) on three Linux link jobs: unit-tests, desktop-core, desktop-e2e-relay. Not added to clippy (check-mode, no link), cross-compile (links inside the cross container), or macOS (ld64 beats mold on Apple Silicon).

• Concurrency group on ci.yml: superseded PR pushes cancel in-flight runs. Mirrors the existing group already on docker.yml. Never cancels main/release pushes so cache seeding is unaffected.

• cross check on PRs, cross build on main (server-cross-compile): catches compile/build-script/proc-macro errors on PRs; link-only musl errors surface on the main push. Saves ~5+ runner-min per Rust PR and removes the future wall-clock cap.

CI — smaller wins

• Flutter pin + cache (mobile job): flutter-version: 3.41.7 matches the hermit pin. Eliminates floating-stable drift against local dev and adds cache: true.

Local dev

• .cargo/config.toml (new): debug = "line-tables-only" for [profile.dev]. Full file/line backtraces, no per-variable debug info — ~20–30% faster local builds and much faster link on macOS. Applies to both the root workspace and desktop/src-tauri via cargo config discovery (cargo walks up from cwd; workspace exclusion does not affect config files). Also shrinks CI test binaries since the test profile inherits dev, superseding the CARGO_PROFILE_TEST_DEBUG=0 approach. If you need lldb variable inspection: CARGO_PROFILE_DEV_DEBUG=2 cargo build ….

• .vscode/settings.json (new): rust-analyzer.cargo.targetDir: true sends RA's cargo check to target/rust-analyzer/ so it no longer invalidates the cargo build artifact cache. Cursor reads .vscode/ too. Updated .gitignore to track this file.

• Remove clippy from lefthook pre-push: both rust-clippy and desktop-tauri-clippy duplicated the CI rust-lint job and cost 2–4 min per push. CI still gates merges. Fast tests stay in pre-push.

Release — parallel platform builds

• New create-release job in release.yml (~30s): derives and validates the version, creates the versioned GitHub release (no assets), and ensures buzz-desktop-latest exists. All four platform jobs (release, release-macos-x64, release-linux, release-windows) now depend only on create-release and run in parallel. Previously they all serialized behind the full ~12-min arm64 build + notarization via needs: release. Expected: ~25 min → ~13–15 min.

Verification

For the profile.ci change specifically: re-run the integration shards a few times to confirm no e2e flakiness at opt-level 1. If Playwright timeouts reappear, bump [profile.ci] to opt-level = 2 — still ~25–35% faster than 3 and well above the opt-0 threshold that triggered #307.

For the release parallelization: test with a -test prerelease tag (routes to --prerelease --latest=false via the version regex) to confirm all four platform jobs start within ~1 min of each other.