fix(pipeline): make run_sift.py invocable by path; skip infra files in entity resolution

[chrisdavidson]

Two defects found while running /epistract:ingest against an FDA product label corpus (Ozempic MAR2022 SPL). Both are invocation/hygiene bugs rather than extraction bugs.

1. run_sift.py cannot be invoked by path

commands/ingest.md documents invoking this script by absolute path. That fails:

$ python3 core/run_sift.py build ./out --domain domains/drug-discovery/domain.yaml
ModuleNotFoundError: No module named 'core'

run_sift.py:21 imports core.domain_resolver at module scope, but running a script by path puts core/ on sys.path[0] — not the project root. The file already inserts the project root, at line 240 inside cmd_dashboard() for examples.workbench.* imports; that is far too late for an import that runs at module load. Until now it only worked with an explicit PYTHONPATH=<repo root>.

Fixed by hoisting the bootstrap above the import, using the idiom already established in core/build_extraction.py:18-22 and core/normalize_extractions.py:43-45.

2. entity_resolution.py parses its own infrastructure files

preprocess_extractions() globs extractions/*.json unfiltered, so it also reads _normalization_report.json. The report has no entities key, so nothing is mutated — but it is parsed on every run and inflates the returned files_processed count by one per infrastructure file.

core/normalize_extractions.py:63,167-170 already solved this for the same directory with _SKIP_PREFIXES = ("_",). entity_resolution.py simply did not follow the convention. Applied the same guard.

Related, and deliberately not fixed here

Every graph build logs:

Failed to load _normalization_report.json: 2 validation errors for DocumentExtraction
  document_id   Field required
  document_path Field required

This originates upstream, not in epistract — sift_kg/graph/builder.py:308-313, where load_extractions() globs *.json and calls DocumentExtraction(**raw) on every match, warning on failure. It cannot be patched from this repo.

The only epistract-side lever would be to stop writing the report into extractions/, but that path is contractual — documented twice in commands/ingest.md, asserted by UT-023 (tests/test_unit.py) and tests/test_e2e.py (x2), and specified in tests/TEST_REQUIREMENTS.md. That is a breaking change and out of scope here.

Change 2 therefore does not silence that warning. Suggested follow-up: an upstream issue against sift-kg to apply the same underscore-skip convention in load_extractions().

Tests

Two regression tests in tests/test_unit.py, both verified failing before the fixes and passing after:

• test_run_sift_importable_by_path_without_pythonpath — subprocess invocation by absolute path with PYTHONPATH scrubbed and cwd outside the repo, so a stray '' on sys.path cannot mask the regression.
• test_preprocess_extractions_skips_underscore_infra_files — one real extraction plus a report; asserts only the former is counted and the report is left byte-identical.

Verification

Check	Result
New tests with fixes reverted	both FAIL
New tests with fixes applied	both PASS
Full unit suite	184 passed, 1 failed
Ruff delta on touched files	25 -> 25 (net zero)

The one failure, tests/test_kg_provenance.py::TestPCCVenueProvenance::test_akka_party_referenced_in_response, fails identically on origin/main — pre-existing and unrelated to this PR. For what it is worth, the assertion looks self-contradictory: it requires the string "akka" while its failure message reads "Response does not mention STA", and the fixture response discusses STA's obligations without ever containing "akka". Worth a separate look.

Test plan

• pytest tests/ -m unit — 184 passed (2 new), 1 pre-existing unrelated failure
• Regression tests confirmed failing without the fixes
• ruff check — no new findings introduced
• python3 core/run_sift.py --list-domains from a cwd outside the repo with no PYTHONPATH
• Reviewer: confirm the sift-kg upstream issue is worth filing

[usathyan]

fix merge conflicts and merge away!