e2e: three TestTUIE2E subtests are red on dev, none new — a teardown race and two waits on a card foot the model does not always draw

[AbirAbbas]

What happened

2026-09-15, dev@e2ae913b7 (and reproduced one commit earlier at 62bc4b7c5), make test-e2e-tui (go test -tags e2e -count=1 -timeout 40m -v -run '^TestTUIE2E$' ./internal/e2e/) on linux/amd64 with the profile's OpenRouter key, model deepseek/deepseek-v4-flash. 14 of 17 subtests pass; these three fail:

--- FAIL: TestTUIE2E/ask_here_end_to_end (36.19s)
    testing.go:1464: TempDir RemoveAll cleanup: unlinkat /tmp/TestTUIE2Eask_here_end_to_end4199179322/001/home/v3: directory not empty
--- FAIL: TestTUIE2E/a_refused_task_proposal_draws_no_schema_sentence (110.41s)
    refusedargs_e2e_test.go:58: waited 1m30s for ["1 yes, set it up · 0 no · c change"] and never saw it.
--- FAIL: TestTUIE2E/space_in_the_task_room_pages_the_card (114.52s)
    tui_e2e_test.go:1928: waited 30s for any of ["enter open its room" " · waiting in this conversation · alt+a"] and saw none.
    tui_e2e_test.go:1960: waited 30s for ["new conversation"] and never saw it.

In all three the law the subtest exists for passed and the failure is beside it:

• ask_here_end_to_end: every screen assertion passed; something under the throwaway CODEAF_HOME is still writing home/v3 when t.TempDir cleans up. Intermittent: 3/3 green on re-run at e2ae913, 1 of 3 red at the parent with the same line.
• a_refused_task_proposal_draws_no_schema_sentence: the log records the screen after the turn carries no repair sentence and Invalid arguments: is never drawn. The model declined to re-propose and ended the turn asking the person (not started · the call was refused, · stuck: propose_task was sent the same wrong argument 2 times, then prose ending "Which way?"), so there was no card and no foot to wait for.
• space_in_the_task_room_pages_the_card: space paged the record exactly as pgdown does passed. The first wait met a third landing shape the test's comment does not accept — roster finished today · 1 folded away / ▸ completed hello.txt file write task, foot enter go to that conversation. The second wait's /history page reads tasks · 1 chat · 1 subtask · $0.0027 but lists only ✓ hello file under finished today: the untitled chat row is counted in the header and not drawn in the 14-row frame. That last part is a product question, not a test one.

Replication

Deterministic (no model). None yet for the first; for the other two the screens above are the evidence. A stub-model rig that ends the turn with a question instead of a card reproduces the second's shape.

Field (real models).

make build
go test -tags e2e -count=1 -timeout 40m -v \
  -run 'TestTUIE2E/(ask_here_end_to_end|a_refused_task_proposal_draws_no_schema_sentence|space_in_the_task_room_pages_the_card)' \
  ./internal/e2e/ > e2e.log 2>&1

Needs a key by any road config.APIKeyAt reads and tmux; about five minutes and a few cents. The two card-foot waits fire on most runs; the teardown race on roughly one in three.

Where

• internal/e2e/tui_e2e_test.go, testAskHere (the TempDir cleanup) and the two waits in testTaskRoomKeepsSpace (search enter open its room and new conversation).
• internal/e2e/refusedargs_e2e_test.go (search 1 yes, set it up).
• The tasks page header count versus its drawn rows: internal/tui3 (search finished today).

The fix

• The teardown: whatever still writes under home/v3 after the surface exits is stopped or awaited before the test returns; a TempDir that cannot be removed is a leak, not noise.
• The two card-foot waits: wait on the fact the subtest is about (the turn ended without the repair sentence; the room was reached by whichever foot the landing drew), never on one spelling of a card foot a model may not draw. Widening the timeout is not a fix.
• The /history header: a row the header counts is a row the page draws, or the header does not count it.

Acceptance

• e2e: the three subtests above green three runs in a row on a clean dev, with -count=1.
• e2e: make test-e2e-tui reports ok for the whole TestTUIE2E.
• Unit: the tasks page's header count equals the rows it can draw for a fixture with one untitled chat and one finished task.
• The manual page that covers the tasks page quotes any new wording, and the change entry's invalidates names what people believed before.