pgw#1514: skeleton.build asks WHY before saying WHAT — a collected tree is not an absent index

[PaulFidika]

Supersedes #1078, which GitHub auto-closed (not merged) when its base 2204-shortwrite was deleted on #1072's merge. Same commit, rebased onto master. Both pins are now on master: ccdb490c (pgw#1511+#1513, the shared helper this calls) and c24636bb (pgw#1518).

skeleton.build reported "carries no model_index.json" about trees that HAVE one. Path.is_file() follows the link, so "this tree never had an index" and "this tree's objects were collected" arrived at the same check as the same False — and the second is a fact about the store that the refusal reported as a fact about the checkpoint. pgw#1308 finding 3, for the fourth time.

Measured, not theorised

A projected tree's manifest pin is the only GC root its objects have. On a real 5.6 GB tensorhub/anima@composed-v3 tree (se#790):

removed pin -> collect_garbage
  -> examined=1211  reachable=0  deleted=1211  bytes_deleted=5,640,566,685
stubs=3   dangling symlinks=14   model_index.json among them

Against that same tree, carries no model_index.json becomes a refusal naming 14 collected entries, saying the bytes must be RE-FETCHED rather than re-pinned, and quoting the false string it replaces so anyone grepping the symptom lands on the cause.

Three decisions worth reviewing

The walk is whole-tree. The index merely dies first, so guarding one file would relocate the wrong message to the next component's config rather than remove it. One walk, before anything is concluded.

No second source of truth — twice. The refusal calls #1072's collected_entries / collected_refusal; an independent implementation of both was written first and thrown away. And the tests fold into tests/test_projected_tree_reading.py rather than adding a module — that file's subject line ("who may, who must refuse, and what they say") already covers this caller; the eager bridge is one reader, the skeleton is the other. Three cases, zero new files. An earlier revision added tests/test_collected_objects_pgw1514.py, which pgw#1362 / DESIGN-RULINGS 4.34b correctly refuses — a filename says when a test was written; the tree needs it to say what it exercises. The guard was run, not routed around.

The original wording survives and is now true whenever reached, with a case asserting it — swallowing the genuine case would trade one wrong message for another.

Red proven — and the first attempt at proving it was vacuous

Recorded because it is how a red proof silently stops proving anything: the first revert used git checkout --, which restores from HEAD — still carrying the fix — and reported a cheerful 10 passed. Reverting to the base branch's skeleton.py instead:

test_skeleton_build_does_not_call_a_collected_tree_index_less   FAILED
test_skeleton_build_walks_the_WHOLE_tree_not_just_the_index     FAILED
2 failed, 8 passed

The third case passes without the fix, correctly — it guards preserved behaviour, not new behaviour.

projected-tree suites          39 passed
ruff · mypy                    clean
incident-naming guard          passes, baseline unchanged (126, and it only shrinks)

Filed by the th#2204 lane on this lane's measurement; owned here per coordinator assignment. Possible visible pause: a re-pin relay outranks this issue — if that arrives mid-review I checkpoint and return.