v0.325.3 - A closing path's pointers are not ids

2026-10-15. Found in the dogfood project the same day, when one diamond was linked to its outcome.

What was wrong

• The validator failed on the closing-path script's own output. derive_closing_path.py --write stores each leaf assumption a diamond waits on under closes_on.inputs, as
  {kind: assumption, id: a-006a-2, ref: "opportunities.yml#opp-006.sol-006a"}. That points at an
  assumption defined in opportunities.yml; it defines nothing in active.yml. The v0.318.1
  exemption skips a pointer only when its ref ends in its own id, the shape of a task input, and
  an assumption input's ref ends in its solution. So when two diamonds' paths named one
  assumption, each counted as a duplicate id. In the dogfood project, linking an L2 to its outcome
  put seven of an L4's assumptions on the L2's path too, and validation failed on all seven, in a
  block nobody writes by hand.
• The prefix check warned of a misfiling for the same reason: an archived and an active
  diamond naming one assumption read as a-030a-NNN defined under two sections.

What changed

• Nothing inside a closes_on block counts as defining an id, for the duplicate count
  (_inside_closes_on) and for the prefix check. Its fired entries, which carry a task id with
  no ref, are covered the same way. A pointer outside closes_on still counts, as the v0.318.1
  test requires.

Considered and not done

• Widening the v0.318.1 rule to any ref into another file. It would also exempt a real entry
  outside closes_on that happens to carry a ref, which that release's test keeps counted.
• Writing the assumption's id at the end of its ref. It changes what ref means for every
  reader of the block, to satisfy one check.

Not covered, and what that costs

• An id somebody hand-writes inside closes_on is not counted either. The block is the
  script's store; the keys it keeps from a hand edit (routes_on_record and the others in
  PRESERVED_KEYS) name tasks and do not define ids.

Tests

• tests/python/test_validate_canvas.py: an assumption on two diamonds' closing paths, and a task
  in both their fired lists, is not a duplicate while a real duplicate diamond id still is; an
  assumption on an archived and an active closing path is not a misfiling while a prefix really
  defined under two sections still warns. Both red on 0.325.2.