Four places the declaration outruns the check

[borisdev]

Not expressiveness gaps — those are closed and enforced by docs/probe_builder_features.py. These are places where a GraphSpec claims something that nothing verifies. All four measured 2026-08-30 against cc59d83.

---

3. object as a declared type opts out of checking, silently — the only real bug here

verdict = VariableSpec("verdict", object)     # examples/ladder/stage9_decision.py uses this
check() -> CLEAN

Declaring object disables check_variable_types for that variable and says nothing. That is 0 FOUND rendering as a pass, which .claude/rules/checks.md is most insistent about: NOT CHECKED and 0 FOUND must never render the same.

Fix: check_variable_types emits NOT CHECKED — 'verdict' is declared as \object`, which accepts anything` and folds it into the existing single summary line. The opt-out becomes visible instead of invisible.

Cost: rung 9's output gains a NOT CHECKED line. That is an improvement — it makes a choice legible.

---

1. The value can be the wrong one. Not fixable by declaring harder.

cite = NodeSpec("cite", inputs=(draft_graph,), outputs=(cited_graph,))

async def cite_that_cites_nothing(ctx) -> CaseGraph:
    return ctx.inputs          # returns the DRAFT, untouched

check() -> CLEAN
run -> pmids=[]                # nothing was cited

draft_graph, cited_graph, verified_graph, case_graph are all one type — which is exactly why VariableSpec names them. But naming makes the wiring checkable, not the behaviour.

No action. This is the boundary between what a spec checks and what an eval measures, and it should be documented as such rather than chased. It is also why a battle metric like CitedEdges is the thing that reads: it counts what the step actually produced.

---

2. check_variable_types reads the annotation, and Python does not enforce it

async def annotated_int_returns_str(ctx) -> int:
    return "not an int"

check() -> CLEAN      run -> 'not an int'

No action beyond documenting it. The check complements a static type checker; it does not replace one. Worth one sentence in the check's docstring so nobody over-trusts it.

---

4. A streaming node's return type is never comparable

An async generator has no useful return annotation, so check_variable_types always reports NOT CHECKED for streams=True nodes.

No action. Already honest — it reports rather than passing silently.

---

Repro

All four in one script; paste into the repo root:

uv run python3 - <<'PY'
# see the issue body above for the four cases
PY

Measured with workflow_workbench at cc59d83, pydantic-graph 2.35.1.