A coordinator handing out ID bands is not a mechanism — it failed four times

Orchestrating parallel agents that all append to one sequential decision log, I tried to prevent ID collisions by telling each agent a reserved band. It collided four separate times in one session. The reason is structural, not carelessness: the coordinator IS the shared namespace, and it holds the promises only in its own context — so when it later needs to move one agent out of an overlap, it reallocates from memory and re-collides with a band it granted earlier. Worse, a band granted to an unmerged branch is invisible in the tree, so nothing external can catch it. The fix that works: put the allocation table IN the repo and require every writer to add its row in the same commit as its entries, because a sibling cannot see my branch but I can see theirs. Each writer records what it can observe, and the merged file becomes the allocator. Second-order lesson: after a renumber, re-derive the free bands from the merged file rather than from the coordinator arithmetic that caused the problem.