[noderaven]

Release 1.6.8: archive integrity, the start check, permits, check reuse and Haiku 5.5.

archive-run re-read the plan and spec from their live paths on every call and replaced any archived copy that differed, so a rerun could leave an archive holding a plan the run never ran. When a run had two launches and only one set of transcripts reached the archive, nothing compared the two. Now archive-run pins the plan and spec to the sha256 the latest setup recorded (the live file, else git show or git cat-file --filters at each recorded setup head) and reports kept, pinned, replaced or a plan_unpinned/spec_unpinned warning. It counts the launches the ledger records against those archived (launches_missing, index key launches_recorded; the audit adds a gap per missing launch). --ledger DIR archives from another ledger, --notes FILE... adds hand-written notes, and a run-owned folder's derive.json and setup.json are archived.

Also in this release:

• A start check after the first failing verify reruns each failing command at the run's start point in a run-owned worktree. When every failing command fails there too, there are no repairs and acceptance says checks_failed_at_start instead of checks_failed (still rejected).
• The dry run lists the tasks a widen path will widen (checks.widening); reference.md says to name shared files rather than folders.
• Pre-flight rulings can permit a file outside a task's Files list; kept permits reach task-checks (--permit) and the ledger, so the task's check is not widened for it.
• run-checks records the setup commands it ran (--record-setup, --record-setup-none), and a resumed integration whose merges made no commit reuses an earlier passing check at the same head (scripts/reuse-checks, checks_reused).
• Haiku 5.5 implements light tasks and runs a docs-only final fix and the start check; one that does not finish reruns on Opus. Reviewers and every other gate stay on Opus. run-report has a haiku tier.
• SKILL.md says to remove the run marker once ledger accept covers every reason.

Built by the parallel-lanes run r168 on 1.6.7 (11 tasks in 4 lanes and a join, 115 min, 31 agents, no fix rounds; 14 final-review findings, all fixed). The Haiku change was built directly after the run. Before the release, the release branch's archive-run was run on two real past runs in a temp copy of the history folder: plan and spec pinned to the recorded versions, launches 3 of 3 and 2 of 2, and one audit's 54 stale-spec gaps gone. Full suite on Windows (Python 3.15): 1862 tests, 1846 pass, 16 skipped, 0 fail.