Retry a timed-out follow-up whose send then fails; fix the gate after #320

[scgopi]

Makes main green after #320 and fixes the defect its exactly-once bookkeeping left behind. Two commits: the gate fixes the Release loop verified (.claude/handoff/gate-fixes-320.patch, applied as-is), then the store fix with tests.

Gate fixes (commit 1)

• DrainWedgeVerificationTests.swift passed deliveryDeadline: last, but GraphStore.init declares it second, right after graph:. Swift matches an argument list in declaration order, so the test target did not compile; the reported messageNode / refreshUsage errors were downstream noise. Moved it to where it is declared.
• ProjectPersistence.swift failed swift-format --strict at 61:10, 62:1, 67:14. Reformatted with the formatter on that file only.

The defect (commit 2)

#320 keeps a timed-out delivery on the queue until its abandoned send reports a verdict. A late success is removed, exactly once — correct. A late failure went through staged(_:), which only writes the memory log and returns a marked copy; the return was discarded in both finishTimedOutDelivery and the drain's fold-back, so the item silently left the live queue. A target that was idle and answering never received text it was owed, while a send that failed before the deadline was retained for retry (the !delivered branch on main). Both failures are now treated the same: staged once, kept in place on the queue, and retried by a drain the verdict itself kicks — a failure learned this late has no command to ride the drain of, so without its own it waited for the next poll, or for good if none came.

The test that was failing

DeliveryWedgeTests.aHungDeliveryStillFreezesEveryOtherLoopsFollowUps asserted delivered would contain the hung one's message, but its own onDeliverMessage closure never appended for the hung node — the hung branch returned true without recording. The last expectation was unreachable whatever the store did, and the store was in fact correct for that path: the send succeeded late, so the message must not be sent again. The closure now records its late success, and the test checks two further drains send nothing more. The new test aTimedOutSendThatThenFailsIsRetriedExactlyOnce covers the late-failure path and fails on main's GraphStore:

✘ aTimedOutSendThatThenFailsIsRetriedExactlyOnce … Expectation failed: (delivered.value → ["[graphcode] for the bystander"]) == ["[graphcode] for the bystander", "[graphcode] for the hung one"]
✘ … Expectation failed: (attempts.value → 1) == 2

Both tests now await the wedged command's return instead of a ten-second settle — the deadline is what returns it — and then refresh, because the bystander's message is queued mid-drain and moves on the pass after. The two suites run in 0.3s instead of 24s.

Not covered here: a send that never reports at all stays parked in pendingDeliveryAttempts, unstaged, until the daemon restarts. Same as #320's main; noted so it is not mistaken for fixed.

Gate

In flight from worktrees/fix-320-lost, private DerivedData; numbers posted below when it finishes.

• make check exit 0 (swiftlint 0 errors, swift-format clean)
• targeted DeliveryWedgeTests + DrainWedgeVerificationTests: exit 0, 6 tests; red on main's GraphStore.swift (exit 65, the new test only)

🤖 Generated with Claude Code

https://claude.ai/code/session_014niBD1D1yQ7F2gjV2K54eB

[scgopi]

Gate on 2fd34769 (= remote head), from worktrees/fix-320-lost, private DerivedData:

Step	Result
xcodebuild test (scheme graphcode)	exit 0 — 1692 tests / 180 suites / 0 failures; 0 ** TEST FAILED **, 0 Failing tests:
graphcode-cli build	exit 0
graphcoded build	exit 0
make check	exit 0 — swiftlint 0 errors, swift-format clean
scripts/cli-smoke.sh	exit 0 on all five verbs, against a fresh swift build of this head (the worktree's .build link points at the main checkout's stale debug binary, so the script's own build was bypassed with a scratch path)
Linux build	pass — https://github.com/scgopi/GraphCode/actions/runs/34087301065/job/101633685121
Red on main	aTimedOutSendThatThenFailsIsRetriedExactlyOnce exit 65 with main's GraphStore.swift (attempts → 1, hung message never re-sent); exit 0 with it restored

Also, #318's PresencePollingTests ran 30 iterations in isolation on this head (-test-iterations 30 -run-tests-until-failure): 300/300 passed.