/we:find-dead-code

Code referenced only from tests is dead: delete it together with its tests. Search production code
only. Source root = the first of app/, src/, lib/ that exists; tests = tests/ or test/.
Search with rg or the repo's code graph (when the instruction files name one, ask it for callers
first), never grep -r: a gitignored graph or cache JSON under the source tree floods the output.

Phases

1. Methods without production callers. For each method in services, CRUD and core modules:
   rg -n "\.<name>\(" "$SRC" -g '*.py'. A repo script for this (find_dead_methods.py or similar)
   comes first. Keep what a framework calls: Pydantic validators, route and signal decorators,
   management commands, middleware hooks, dynamic dispatch (getattr, dict lookup), anything
   annotated as an intentional interface.
2. 0 % coverage files, only when pytest-cov is installed:
   pytest "$TEST"/unit --cov="$SRC" --cov-report=term -q --tb=no | rg ' 0\.00?%'. Confirm with
   rg that nothing in $SRC imports the file; zero importers → delete.
3. Test-only modules. Per module, count importers in $SRC (excluding itself) and in $TEST:
   zero production importers and at least one test importer → dead. A patch("pkg.mod.fn") target
   is a mock, not a caller.
4. vulture at --min-confidence 80, migrations excluded; never lower, the rest is noise. Skip
   TypedDict fields, enum members used only in annotations and TYPE_CHECKING imports.
5. Clean up and gate. rg -l "<deleted symbol>" "$TEST", fix the imports, then the repo's
   formatter and linter and the affected unit tests by name.
6. Repeat phase 1 after every deletion batch until it comes back empty: a deletion exposes the
   next layer.

When the repo carries a dead-code gate (for example a scripts/check-dead-* script), run it last.