ci: seed test-e2e's warm DerivedData from main every 6 hours - #14082

A focused test-e2e.yml dispatch should compile only what changed. #14016 made that possible: a build starts from DerivedData that a main-contained dispatch published. But nothing publishes that DerivedData regularly. Focused dispatches test pull request revisions, which never publish, and the seeds so far came from hand dispatches, whose artifacts expire after 3 days. When no seed exists, the build logs "no-main-derived-data" and compiles the whole app.

Changes

• e2e-derived-data-seed.yml (new). Every 6 hours it dispatches test-e2e.yml on main's HEAD with cmuxTests/CmuxPopoverGroupTests (36 ms), unless that SHA already has a seed run that finished or is still running. It dispatches rather than calls, so the run has the identity e2e_warm_derived_data.py trusts: path test-e2e.yml, event workflow_dispatch, branch main. This mirrors how ci-main-full-suite.yml dispatches ci.yml.
• Seed runs always compile. Only a compile publishes DerivedData. A seed run could adopt an exact product instead, because product identity ignores .github/ and tests/, and a focused dispatch of the pre-merge PR head is a valid producer. It would then publish nothing. test-e2e.yml now skips product reuse when dispatch_id starts with e2e-derived-data-seed-.
• Not in this PR: clearing a half-adopted DerivedData after the adopt step times out. Main landed that while this was open, and the merge takes main's version.

Validation

• All 133 commands in ci-guards.yml run locally, including those inside run: | blocks. bun test/claude-environment.test.ts fails, as it does on main. test_local_build_cache_preflight.py failed once under sweep load and passes 22 of 22 when run alone; its tests are deadline-based.
• actionlint is clean.
• An independent review found the same warm key for blacksmith-6vcpu-macos-26 and the 12 vCPU pool that ci: route focused tests to one front door, and half of them to 12 vCPU #14067 routes odd SHAs to. Both run image 26-20260916 with Xcode 26.6 and /Users/runner/_work/cmux/cmux, and both computed f9243ea2… (runs 35939305209 and 35939623231). An image rollout that reaches one pool first would split the key until the other catches up.
• The first seed was dispatched by hand: run 35941008193, on main a3b7014. Its dispatch id isn't a seed id, so the first scheduled tick may run one more build.

Not yet shown

• ci(e2e): start builds from main's DerivedData so test-only changes skip the app compile #14016's open question: whether Xcode 26 rebuilds only the changed test file from adopted DerivedData. The first focused dispatch after a seed answers it; the build summary reports the adoption and how many inputs differed.
• Changing test-e2e.yml changes the E2E product recipe fingerprint, so exact-product reuse misses once after merge.
• Cost: one e2e build every 6 hours on free Blacksmith, skipped when main hasn't moved.

— Camera g1 🛠️

🤖 Generated with Claude Code

Summary by CodeRabbit

• Chores
  • End-to-end testing now includes scheduled runs that prepare derived data, alongside the existing manual test runs. These scheduled runs use the latest code on the main branch and avoid reusing compiled products for the seed run. No changes to app functionality are included.