[lavaman131]

Summary

e2e/document-entry.e2e.ts › "a parent rename while its child closes publishes the parent as the visited document" no longer shares its account with the rest of the suite. It signs in as ana-<its room id> instead of ana, so no other test's page can overwrite the last visited document it polls for. The product is unchanged.

Cause

The test ends by polling GET /api/navigation until lastDocumentId is the parent. That value is one row per account (user_navigation.last_document_id), and PATCH /api/navigation replaces it, last write wins, from any page of the account. The test signed in as ana, as do the file's other three tests and eleven other spec files, and the tests of one run share one server and one database. A page another test opens as ana records a visit of its own, and if it lands between the parent's visit and the poll's first read, the poll never sees the parent: it reads the other test's document until the 10 s timeout.

Evidence, from an unfixed checkout of main:

• bun run e2e e2e/document-entry.e2e.ts (4 workers) failed this test at the lastDocumentId poll in 2 of 2 runs, with Received a document other than room, exactly as in the CI log (job 126809564779, shard 11/12, all three attempts). --repeat-each=30 failed it 4 of 30; the file's other three tests passed 90 of 90.

• The failing run's Playwright trace has exactly two PATCH /api/navigation requests from the page: its child, then room. Navigation reads between them returned the child, then the parent; every read of the poll afterwards returned a document the page never visited.

• Temporary server logging of each committed visit (not in the diff) for one failing run, all for ana on the same server:

  Time (s)	Document visited	Test
  24.116	1400ef89 (child of e814ebe3)	this test, child opened
  24.176	b6819b74	"canonical title changes preserve a pending child request…"
  24.227	e814ebe3 (the parent)	this test, after Back: the visit it waits for
  24.236	4780ae72	"a delayed direct child cannot reopen…"
  24.415, 24.434	b6819b74 (renamed)	"canonical title changes…"
  25.195	d3bb39d8 ("Source review a2d1b0f5")	"a direct child opens its parent from one read and recovers…", after Try again

  The poll expected e814ebe3 and received d3bb39d8.

The test's own visits are recorded in the right order (the client chains them through visitTail), and a person's last visit being whichever tab visited last is the intended behaviour, so there is no product race to fix. It is an isolation defect in the test: it asserts on account-wide state that its neighbours write.

Fix

One line in the test: authenticate(page, `ana-${room}`, baseURL!), with a comment saying why it is not the shared ana. document-creation and hosted already use a handle per test for the same reason. The assertions are untouched, so the test still catches a wrong last visit: with the server temporarily made to ignore visits to parent documents, it fails 3 of 3 at the same poll.

Verification

All on this branch's head unless noted.

• Red before the fix (above): 2 of 2 runs of the file and 4 of 30 repeats failed at the poll.
• E2E_SKIP_BUILD=1 bun run e2e e2e/document-entry.e2e.ts --repeat-each=30: 120 of 120 pass (4 workers, production client built with the child-motion toggle).
• The same under taskset -c 0,1: 120 of 120 pass, 6.5 min.
• Shard 11/12 (the one that holds document-entry; shard assignment from e2e/shard-balance.ts and shard-timings.json): CI=true bun run e2e --shard=11/12 --workers=2 --reporter=github,blob,list,./e2e/shard-balance.ts. The first attempt had 67 passes and 2 machines.e2e.ts failures at git commit in their fixture repository, because this machine's global git config signs commits with a key defined only under a work directory (CI has none). With GIT_CONFIG_GLOBAL=/dev/null: 69 of 69 pass, none retried.
• Mutation check as above.
• bun run fix changes nothing beyond the edit (72 existing lint warnings, none in the file); bun run types and bun run ci pass.
• bun test was not run: no product code changed. bun run db:up could not start because port 5432 was taken by another container; bun run e2e starts its own four databases.

Not reproduced: CI's three failed attempts in a row. The mechanism explains them (each attempt has the same neighbour worker opening documents as ana), but I only reproduced the failure locally at the rates above.

Fixes #783

Assistant-model: Claude Sonnet 5.5
Assistant-workflow: goal (run 4545dc40-39f5-4e69-8642-76e0fee3e70d)
Assistant-duration: 28m converged, estimated 1h
Assistant-verification: red before fix passed: E2E_SKIP_BUILD=1 bun run e2e e2e/document-entry.e2e.ts on main failed the visit test at its lastDocumentId poll in 2 of 2 runs, and --repeat-each=30 failed it 4 of 30 while the file's other three tests passed 90 of 90, each time with another test's document as lastDocumentId
Assistant-verification: server visit log passed: with temporary logging of each committed PATCH /api/navigation (reverted), the test's child then parent visits committed in order and a later visit by the file's first test became the last record for ana
Assistant-verification: E2E_SKIP_BUILD=1 bun run e2e e2e/document-entry.e2e.ts --repeat-each=30 passed after the change: 120 of 120 with 4 workers, against 4 failures in 120 before it
Assistant-verification: taskset -c 0,1 bun run e2e e2e/document-entry.e2e.ts --repeat-each=30 passed: 120 of 120 pinned to two CPU cores
Assistant-verification: CI=true bun run e2e --shard=11/12 --workers=2 --reporter=github,blob,list,./e2e/shard-balance.ts first attempt failed: 67 passed, and 2 machines.e2e.ts tests failed at git commit in their fixture repository because the local global git config signs commits with a key defined only under a work directory; the document-entry tests passed
Assistant-verification: GIT_CONFIG_GLOBAL=/dev/null CI=true bun run e2e --shard=11/12 --workers=2 --reporter=github,blob,list,./e2e/shard-balance.ts passed: the 69 tests of the shard that holds document-entry, none retried
Assistant-verification: mutation check passed: with the server ignoring visits to parent documents, the visit test failed 3 of 3 at the same poll; restored, it passes
Assistant-verification: bun run fix passed: no change beyond this edit, 72 existing lint warnings, none in the file
Assistant-verification: bun run types passed: whole workspace and e2e
Assistant-verification: bun run ci passed: dprint, oxlint (72 existing warnings, 0 errors), token and design checks
Assistant-verification: bun test unavailable: not run, no product code changed
Assistant-verification: bun run db:up unavailable: port 5432 was taken by another container, and bun run e2e starts its own four databases
Co-authored-by: Alex Lavaee lavaman131@github.com