Non-normative. Directions outside the contract, recorded so their absence is deliberate rather than overlooked. Nothing here binds the design.
Two agents taking the same Step is outside the contract. The plausible direction is a serializing owner rather than claim leases: leases require a clock and a reaper, and produce ambiguous states when a holder dies mid-work.
Divergence does not cover this. Divergence reconciles concurrent revision of intent, and the availability tradeoff (CMP-T01) that makes it tolerable does not obviously transfer — two agents doing the same work twice is waste, not a disagreement with two valid sides.
Divergence resolves by authorship, and a reconciliation can itself diverge, since the model has no serialization point. Two directions are plausible.
A deterministic reconciliation function for the common cases — disjoint step additions being the dominant one — would let two machines independently produce byte-identical results, converging without coordination and preserving CMP-T01. An elected serializer per Plan would also work, at the cost of that tradeoff.
The first is preferable, and depends on distinguishing genuine disagreement from staleness: a machine that revised from an older version only because replication lagged did not actually disagree, and that case is mechanically reconcilable.
Constrained by decision 0003 Amendment 3: no-delete replication means removal is unavailable, so compaction cannot work by deleting, and a tombstone leaves files in the scan path.
The plausible direction is a summary version subsuming a range of lineage. It does not follow from the index: the index makes re-reading cheap and does nothing about how many files exist or how many are scanned to discover them.
A summary version is harder here than under an inert format, because a later version references its parent and evaluating it evaluates the chain. A summary that is not referenced saves nothing, and one that is referenced changes what the lineage says.
Rendering a Plan into an external issue tracker for human or cross-organization visibility. Compass remains the authority and the projection is one-way. Bidirectional sync is explicitly not a direction: it reintroduces the dual-authority problem the design exists to avoid.
There is no path from existing trackers or checklists. A Plan's value here is its rationale chain, and imported work has none, so an import yields a Plan whose lineage begins with a single uninformative entry. Honest, but thin enough that it is worth doing only if adoption depends on it.