phase2: the inventory's disposition verdict runs in no gate, and the committed phase2-inventory.txt is stale (#1758) #1804

Found by the #1758 R1.1 round-2 blind critic (PR #1776), 2026-10-15, on branch phase2-numbers at cdfb128. Both files below exist only on that branch.

Design 2.5 and 2.7 (R1.1, row 2) say a slot classifier with no reviewed disposition is caught because "the inventory fails on an unreviewed entry". The inventory does fail. But no gate runs that verdict:

• tools/phase2_check.sh --through R1.1 runs tokens, int-source, slot-test, corpus-diff and smi-trap. None of them runs tools/phase2_inventory.py's verdict.
• The slot test's runner calls phase2_inventory.py --slot-population, which returns (line 832) before the disposition check (line 943). It lists only function entries whose first parameter is an EigsSlot.
• No workflow, precheck row or suite section runs the full inventory.

Reproduce

Append a new classifier to a copy of src/vm.c:

int critic_new_reader(Value *v) {
    EigsSlot s = slot_from_heap(v);
    return slot_is_bool(s) ? 1 : (slot_is_null(s) ? 2 : 0);
}

Results on the copy:

• python3 tools/phase2_inventory.py --no-write exits 1:

  FAIL: slot classifier `vm.c:critic_new_reader` has no disposition row
  slot classifiers: D1=56 D2=19 D3=25 T=67 dispositions=99 missing=1 stale=0
  phase2-inventory: FAIL
  

• --slot-population still prints slot-population: entries=33, so the slot test is unchanged.
• The token gate (tokens: rounds<=R1.1 tokens=5 files=89 hits=0) and the int-source check both pass.

So phase2_check.sh --through R1.1 (and R1.2, whose 464 forced sites are where new classifiers appear) passes with an unreviewed classifier.

The committed generated file is stale too. docs/design/phase2-inventory.txt says "generated by tools/phase2_inventory.py; do not edit", but at cdfb128 it differs from a fresh --out in 753 lines:

• examined=1818 (fresh: 1826);
• dispositions=98 (fresh: 99);
• NUM_CMP, INT_BINOP_R, CASE(NEG) and CASE(BNOT) are still recorded guarded (fresh: unwired, the R1.1 fix round's change);
• #define SLOT_SMI_TRAP is missing.

phase2_check.py's plan checks read their ## Numeric arms from that committed file.

Done when

• phase2_check.sh --through R1.1 and --through R1.2 run the inventory's verdict as a mechanical gate, and it fails on phase2-inventory: FAIL. The planted critic_new_reader above turns that gate red and names the function.
• At the branch head, docs/design/phase2-inventory.txt equals a fresh tools/phase2_inventory.py --out byte for byte, and the same gate fails when they differ. Alternatively, the file is no longer committed and the plan checks derive the numeric arms themselves.