[noderaven]

Release 1.6.5: join waves.

Until now the join ran its tasks one after another on the feature branch after the lanes were merged; on d1a it was 274 of 630 minutes, all on the critical path. Now, in the launch that starts the join, tasks whose depends_on allow it run side by side, each on its own branch in a lane worktree, and each wave is merged back by recorded sha (scripts/merge-wave) before the next one starts. A join task without depends_on waits for every earlier one, so older manifests keep a serial join. On d1a's join the waves are 1-2-4 (task-slots 17 to 13).

Also in this release:

• One check after the last wave with more than one task: one fix, a re-review and a recheck; an autonomous run hands a remaining failure to the final review and verify, a supervised run stops.
• A merge conflict goes to a resolver (Opus) in an autonomous run; a supervised run stops and gives the exact git merge --no-ff command to run by hand.
• Pre-flight sees the waves and can move a task to a later wave (preflight.wave_moves).
• A relaunch is safe whatever the join's state: setup returns lane worktrees to their lane branches (saving uncommitted changes under the abandoned refs), a recovery merge brings in the finished and reviewed tasks, and the rest of the join runs serially from that tip; a merged task whose approval went stale is reviewed again in place.
• PL_PHASE for project scripts (run-checks --phase, task-checks --phase), so a check script can tell a join task in a lane worktree from a lane's own work.
• The ledger takes concurrent writers (the lane file's OS lock; 4 of 400 concurrent appends were lost on Windows without it) and skips a half-written last line.
• Dry run, owners, worth-it (--join-deps), the back-test (--waves) and autopsy know the waves.
• Lane ids may no longer start with u-.
• CI: a job pinned to Python 3.15, no Node 22 job; the test suite needs Node 24 or later. Two test-only faults CI found are fixed: the platform test faked Windows before importing the new lock module, and the waves tests used load_module, which Python 3.15 removed.

Testing: built directly, test first (15 plan tasks); a builder check and four adversarial checks of the design by fresh agents testing the git mechanics in scratch repos; an end-to-end test on real git (a two-task wave, then a stop and a relaunch); a final review by a fresh reviewer (one critical and three important findings fixed test-first, seven minors deferred). Back-test of 31 archived runs: same verdicts at threshold 1 with or without waves. Full suite on Windows (Python 3.14 and 3.15): 1685 tests, 1669 pass, 16 skipped, 0 fail; CI passes on Ubuntu (Node 24, and Node 24 with Python 3.15), macOS (Node 24, Python 3.15) and Windows (Node 24, Python 3.12). Not run: a validation run on a real project.