Evidence: the red run (T0)

The new and rewritten assertions, run against unfixed code. This is the counter-evidence for the claim a bug fix is easiest to fake: two of these tests previously asserted the 400 behaviour, and rewriting a test is only honest if the rewritten test is shown failing first.

Captured 2026-08-16, on claude/github-issue-238-cwdgone at the commit before task 3. Absolute paths are redacted to /Users/… and <tmp>; nothing else is edited.

Python — cli/tests/test_core_graphs.py, cli/tests/test_api_routers_integration.py

$ uv run pytest cli/tests/test_core_graphs.py cli/tests/test_api_routers_integration.py
=========================== short test summary info ============================
FAILED cli/tests/test_core_graphs.py::test_repo_resolves_agrees_with_resolve_repo
FAILED cli/tests/test_core_graphs.py::test_check_answers_a_vanished_checkout_instead_of_raising
FAILED cli/tests/test_core_graphs.py::test_the_unknown_position_answer_is_not_a_filesystem_oracle
FAILED cli/tests/test_api_routers_integration.py::test_graph_check_answers_a_checkout_that_has_been_deleted
========================= 4 failed, 19 passed in 1.21s =========================

The four failures, and why each is the right kind of red:

Test	Failure against unfixed code
test_repo_resolves_agrees_with_resolve_repo	AttributeError — graphs.repo_resolves does not exist yet.
test_check_answers_a_vanished_checkout_instead_of_raising	ValueError: repo path is not a directory: <tmp>/nope — the behaviour being removed, raised from the call that should have returned.
test_the_unknown_position_answer_is_not_a_filesystem_oracle	Same ValueError, raised before the body it asserts on exists.
test_graph_check_answers_a_checkout_that_has_been_deleted	assert 400 == 200 — the exact status code the ticket was opened about.

The route-layer failure in full, since it is the one that mirrors the reported symptom:

        response = client.post(
            "/api/v1/graph/check",
            json={"repo": str(tmp_path / "nope"), "workItem": "issue-1"},
        )
>       assert response.status_code == 200
E       assert 400 == 200
E        +  where 400 = <Response [400 Bad Request]>.status_code

cli/tests/test_api_routers_integration.py:101: AssertionError

The two tests that assert unchanged behaviour pass here as they must — test_a_resolving_repo_keeps_exactly_the_keys_it_always_had and test_graph_check_says_nothing_new_about_a_checkout_that_is_there are green before the change and must stay green after it (R2.2).

UI — ui/src/state/useControlPlane.test.ts

$ cd ui && bun run test
- "outer": {},
+ "outer": {
+   "github:acme/widgets#7": {
+     "currentNode": "",
+     "nodes": [],
+     "ok": false,
+     "repoResolved": false,
+     "workItem": "issue-7",
+   },
+ },

 ❯ src/state/useControlPlane.test.ts:47:21
     45|     const reports = await fetchGraphs(api, [WORK_ITEM], [SESSION], new…
     46| 
     47|     expect(reports).toEqual({ outer: {}, inner: {} });
       |                     ^

 Test Files  1 failed | 7 passed (8)
      Tests  1 failed | 105 passed (106)

One of the two cases is red: fetchGraphs stores the unknown-position answer instead of dropping it, which is precisely what would replace railFromFrozen with an empty rail. The sibling case — a normal answer is stored — passes before the change and must keep passing after it.

bun run typecheck is also failing at this point, because GraphStatus has no repoResolved field yet. That is task 4's, not a second defect.