Remove nested turbo from package build scripts

[Vadman97]

Summary

• Replace \"build\": \"yarn turbo run prepublish\" in each package with the command the nested turbo was going to run (e.g. tsc -noEmit && vite build), matching the existing pattern in packages/utils.
• The outer turbo invocation (e.g. turbo run build from a consumer repo) already walks the dep graph via ^build, so the nested layer was redundant — all it did was re-enter turbo from inside a workspace already being scheduled.

Why

On Linux, yarn build of a consumer project that depends on these packages (we hit this in launchdarkly/observability-sdk) would hang indefinitely with no subprocess activity. Every turbo process stuck in futex_wait_queue. No vite, no tsc ever spawned.

Root cause: each nested yarn turbo run prepublish re-runs turbo's root discovery from the package's own CWD, finds rrweb/package.json (with workspaces), and computes a repo-root hash different from the outer turbo's root hash. That means it tries to start its own turbo daemon — separate inotify instance, separate Unix socket under /tmp/turbod/<hash>/. When 10+ rrweb packages build in parallel, several nested turbos race to create the same daemon. On Linux, losing that race leaves you with a half-open socket connection to a daemon that's already been replaced, and you sit in futex_wait forever. macOS rarely hits this because kqueue/fsevents and BSD socket semantics resolve the race faster.

Observed symptom when stuck:

/tmp/turbod/
├── 2948c6004e604b3c/  # rrweb-snapshot CWD
├── 8be077b6d3282621/  # rrdom CWD
├── 973b7cc49da23c74/  # outer turbo root
└── e9931d89ef6fbc99/  # types CWD

Four daemons for one build = four root hashes = turbo thinks each package is its own root.

Additionally, most of these scripts were yarn turbo run prepublish without a -F filter, so each nested turbo tried to schedule prepublish across the entire rrweb workspace again — a graph the outer turbo had already walked.

After

Only one turbo daemon exists (the one the outer run starts). yarn build:sdk in observability-sdk goes from "stuck, need to cancel and retry 3× on Linux" to ~58s deterministic.

Semantics are preserved: each new build script is the verbatim body of the same package's prepublish script. prepublish is left in place since it's a yarn/npm publish lifecycle name.

Test plan

• yarn build in observability-sdk on Linux — completes cleanly (18/18 tasks, no hang)
• Verify CI still passes here
• macOS sanity check (should be unaffected since it already completed; just faster now because of one fewer layer)

🤖 Generated with Claude Code