Fixes the false negative in reference extraction reported against #458.

The defect

Both validators answered "is this name the template's own?" with one template-wide set of every bound name. That is right for a name used only where it is bound, and wrong everywhere else — binding a name anywhere silenced it everywhere:

{{ ghost.done }}                <- undefined; reported nothing
{% for ghost in rows %}
  {{ ghost.name }}              <- the binding that silenced it
{% endfor %}

The reference outside the loop is exactly the typo a validator exists to catch, and adding an unrelated loop elsewhere in the file made it vanish. The trade was backwards: the false positive this replaced was loud, a false negative is silent.

The fix

core/template_scope.py tracks scope the way Jinja does. A binding reaches the body of the construct that introduces it and nothing outside it; {% set %} reaches the statements that follow it in the same block. {% if %} opens no scope, because Jinja gives it none — treating its {% set %} as reaching the rest of the block can only suppress a report, never invent one.

Identity, not spelling: shadowed_name_nodes returns the id() of each Name node that refers to a local binding, so one template can hold both a shadowed and an unshadowed use of the same word.

Both validators ask it, so they cannot disagree about what a template defines.

All seven cases from the review are covered, plus: the iterable evaluated outside its own loop, {% for %}'s filter test, the {% else %} branch, loop itself, tuple targets, {% set %} reading its own prior value, {% call %} arguments, and macro defaults.

Verification

• Reverting to the template-wide bound set fails 20 tests.
• Blocking suite: 772 passed, 13 skipped, 0 failed.
• Catalogue: 50/117, unchanged — the fix reports strictly more, and no validating example depended on the over-broad suppression.
• Lint clean.

Two things worth flagging

The end-to-end guard initially passed under the very mutation it existed to catch. Two reasons: the probe pipeline's rows is undeclared, so it is refused either way; and the template validator finds undeclared names with jinja2.meta, which is scope-aware already and reported ghost on its own. Only the data-flow validator was blind. The assertion now targets its specific phrasing, and fails under the mutation as it should.

Three dead imports removed — jinja2.meta, TemplateSyntaxError and Union in data_flow_validator.py. meta reports only top-level undeclared names, which is why the chain and subscript forms had to be recovered from raw text; walking the AST replaced both, and these have had no caller since #458.

🤖 Generated with Claude Code