[DanielNetzer]

Closes #220.

This is a test-only change across molecule, harness, process and cli. No production file changes, and every shipped default is untouched.

• Readiness markers instead of wall-clock waits: a new cancelWhenStarted helper cancels a run only after the stub has written its "started" marker. Interruption tests no longer race a 10 s watchdog, and a stub that never starts fails explicitly.
• Marker-driven expiry in process: expireOnMarker and markedShell let tests that need output and then a timeout expire only after the output provably exists. Silent-stall tests still wait out a real stallTimeout, because there the deadline is the behavior under test.
• Success backstops: tests that only need a run to finish use a named 2-minute successBackstop instead of a 10–30 s budget. That replaces the 10 s budget in TestNativeBindingPassesAllowlistToClaudeArgv and the 30 s budgets in analysis_correction_test.go. The shared run fixture's check timeout goes from 5 s to 120 s.
• TestHarnessTimeoutPlanValueIsHonouredAndRecorded: the control run now uses a stub that never finishes, so it still proves the Formula timeout fires. The plan-extended run gets minutes of headroom over its 11 s harness, and the test asserts that headroom.

Built by Forgecell Molecule mol-220-05bba754ffa48566e90b815a (plan revision 2 added the three extra flaky files the coder found). On the candidate tree, the load acceptance go test -count=5 -p 8 ./internal/molecule/... ./internal/harness/... ./internal/process/... passed, along with the 10× repeat of the three named tests, the cli/install packages and the shipped-default assertions.

Hand-delivered: the Lab's full-suite baseline failed only on Go's 10-minute timeout in the unrelated cli and evaluation packages, which #222/#224 plus the Formula refresh address. Rebased onto main (1f57972), an uncached clean-shell go test -count=1 -timeout 25m ./... passes, and gofmt and vet are clean.

🤖 Generated with Claude Code