[AbirAbbas]

Fixes the test failures still red on dev after #2086. Two of them were real bugs in taking a chat from another computer.

What changes

• A displaced window keeps its unsent turns (code). When an ownership check (CheckLease, from the directory watch or before a tool call) found another device holding the chat, the window became a viewer at once and never told the batcher. So unsent turns were not kept as a branch then. Now:
  • The hold is retired at once, so Gate refuses new work and the send receipt is cleared.
  • A seal already past Gate is allowed to note its turn.
  • Batcher.Displaced resolves the pending work under flushMu.
  • Only then does the window say who has the chat, exactly once.
  • A publish refused with a stale fence goes the same way (OnLost now retires the hold, and resolve announces), so the window no longer says another device continued this chat and then <device> continued this chat a moment later.
  • A direct refusal(ErrFenceStale) still makes the window a viewer at once.
• A take no longer trips over the sender's old upload note (code). A snapshot carries the sender's .cell/published.json, which names the head published before it. The receiver never fetched that head, so the next take failed with snapshot … is not held in this store. Materialize now records the head it received as published. That head is on the service by definition.
• R7 flake (test). personalLeader returned as soon as Status answered, while other peers' replies were still on the member socket. R23 then counted those frames as traffic started after the quiet decision, about 1 run in 20. It now waits for the member to go idle. The assertion is unchanged.
• TestPairJourney. The add-device pitch now comes through the tuiwords table instead of a retyped needle. The home tip is respelled space space brings you back to home, because the journey's vocabulary law forbids "take". The manual pages that quote it are updated.

Proof

These two fail on current dev with the engine staged, and pass on the branch:

Test	Failure on dev
TestTwoHomesL12DeathBranches	A's branch row = {…}; want 3 turns from tower
TestPublicTakeRestoresEveryWorkspaceSetRoot	snapshot … is not held in this store

CI does not stage the engine, so these real-engine tests do not run there.

• New: TestStalePublishRefusalMakesViewerAndNotifiesOnce, a real stale publish across two homes. The window ends as a viewer, and the notices are exactly [kept edits, superseded].
• TestAStaleDriveRefusalImmediatelyMakesTheWindowAViewer now also checks that two refusals give one notice.
• #2086's versions of TestDriveSideSupersededStopsTools, TestDriveSidePublishesOverTheDefaultDaemon and TestTwoHomesTakeARunningChatAtOnce pass unchanged.
• R7 at -count=100 twice in a row (lane). Locally, R23 did not recur in 50 more runs. Two runs hit a client timeout talking to the local cluster during setup, on a box at load average 8 with other suites running.
• Focused internal/syncsetup, internal/cellsync, internal/cellstore, internal/tui3 tips, internal/manual, untagged internal/e2e, go vet -tags e2e ./internal/e2e/, and make test-laws.

🤖 Generated with Claude Code