[JacobStephens2]

Summary

The Base fix's rules now reach the outside world only through a private Outside seam. There are two adapters: the production one over the Launch directory and GitHub, and a scripted fake used in the module's own tests. The public interface is unchanged.

 BaseFix::fix(launch, issue, …)
-  launch.lock / github::open_issues_labelled / Mark::of / github::create_issue
-  Running::mark / child_run::start+wait / poll::until(issue_is_open) / progress::step
+  fix_through(&mut LaunchAndGitHub { launch, issue }, issue, …)
+    outside.look / open_fixes / mark / create_issue / mark_running
+    outside.run_base_fix / merged
+    wait_on: loop { has_ended (mark first); issue_is_open; pause }
+    open_fix_covering(listed issues, …)   # pure

• poll::pause() is new: a public sleep of one poll interval, which poll::until now uses as well.
• Left detection, meaning the mark file is there and its lock is free on first open, stays in the production adapter. The rule that a Left issue is started again is in the core.
• One change you can see: waiting on Base fix #n is now printed just before the child Run starts. So if the start fails, the line comes before the error instead of not at all.

Evidence

• Before: none of the Base fix's rules had a unit test. Each was checked only end to end, through the fake gh and claude, in 20 tests in tests/base_fix.rs.
  After: 19 unit tests in src/base_fix.rs check the order of Did events and the exact error text. For example:

  covering issue #8 marked Running, issue open → open → closed
    did: Look, ListFixes, ReadMark(8), LookReleased, "waiting on Base fix #8, already open: …",
         ReadEnded(8), ReadIssue(8), Pause, ReadEnded(8), ReadIssue(8), "Base fix #8 closed; …"
  

  The 8 end-to-end tests that only checked a rule were deleted. The 12 that need real processes, locks or the whole Run are kept. cargo test passes in full and cargo clippy --all-targets -- -D warnings is clean.

Merge Danger

Door: two-way

This only changes tests and the internal structure, so a revert undoes it cleanly.

Blast Radius: Base fix

If the production adapter differed from the old code, the Runs that take a Base fix would be affected. The kept end-to-end tests run the happy path, a Left mark restarted across processes, and the Spec run's shared Base fix through the real adapter.

Unaddressed findings

Standards

• Rename Outside::step to progress, as in the Repair loop: skipped, because the issue's seam shape names it step.
• Data clump issue, pr_url, base, base_commit, failed: skipped. It was there before this change, and bundling it would change fix's public signature, which the issue keeps.
• Outside has 11 methods, and create_issue's labels parameter always gets the same value: skipped, because both match the issue's seam shape and its decision to keep labels in the core so the tests can check them.

Spec

• The fake's open_fixes and marks are fixed Vecs, not per-call VecDeques: skipped. A Run lists open fixes and reads each mark once, so per-call scripting would add nothing. The answers that do vary per call (ended, open, children) are VecDeques.

Closes #378

🤖 Generated with Claude Code