[squid-protocol]

Summary

Part of the Level 2 call-graph work (epic #3265).

Problem. A classification of every pyan3 edge the resolver doesn't link on language-crucible's Python repos found decorators to be the largest real gap: 641 edges, 18% of pyan's total.

• 431 of those are decorator calls at module level that were recorded nowhere. Python (Mode C) sliced only defs, so the code a module runs at import time never entered the call graph. That includes app = FastAPI(), @app.post("/items"), setup_logging() and if __name__ == "__main__": main().
• For a web app, that's the route registration and wiring: how a request reaches a handler.

Fix.

• Slicer (detector.py, _module_level_unit): for Python and embedded_python (top-level module_level_unit profile flag, not a rules pattern, per orphaned_logic reports 100% dead code for five languages whose invocation model never names the callee — the family #2727 closed without covering #2806), the Mode C slicer emits Mode D's existing synthetic __global_context__ bucket.

  • It holds the file with every sliced unit's span blanked (newlines kept), and each class Name header's name blanked, since a class statement isn't a call.
  • Calls, qualifiers and receiver types are scanned as usual, so @app.post(...) after app = FastAPI() resolves through Call resolver: typed step — resolve x.method() when the function shows x's class (Python) #3693's typed step to FastAPI.post.

• Calls-only. The unit has magnitude/impact 0 and a calls_only flag. That flag is honoured by everything that sizes, counts or lists functions, so no file's size or classification moves because of it:

  • file mass (signal_processor);
  • the per-file "Function Analysis" report and the function-impact forensics;
  • the file-archetype log_function_count;
  • the statistical auditor's named-structure retention;
  • the folder function count.

  Without this, the first cut moved 3,688 golden-master values. Most of that came from files with no functions losing the fallback magnitude estimate, and from every Python file gaining one "function".

• Delta scans. Synthetic units have no function_data row (functions_found counts the slicer's synthetic __global_context__ buckets, inflating every per-function descriptor for 5 languages #2691), so the state rehydrator could not restore them, and an incremental scan would have lost these calls. They now also go into a new per-file synthetic_unit_data child table (FK cascade, with the calls_only flag), which the rehydrator restores. Mode D's existing buckets are persisted and restored the same way.

Type of change

• Bug fix
• Parsing or engine logic (detector.py, python profiles, signal_processor.py, archetype_classifier.py, statistical_auditor.py, record_keeper.py, audit_recorder.py, state_rehydrator.py)

Measured

Module-level call pairs now recorded, per language-crucible Python repo (confident = scoped/unique):

repo	pairs	confident	steps
fastapi	417	416	typed 193, import 193, file 30
numpy	32	29	file 29, unseen 3
wtfpython	4	4
cython	3	2
twisted	2	2
airflow	0	0	the excerpted files have no top-level calls

call_graph_resolution.py python (pyan3) is unchanged by design: 99.4% precision, 51.3% recall, 96.9% resolution recall. pyan has no module node, and it attributes a decorator to the decorated function, so these links are invisible to that metric. call_graph_accuracy.py --ci and tree_sitter_accuracy_audit.py --ci --all (30 languages) both pass: the synthetic unit is never counted as a function.

Golden masters

Both fixtures re-blessed: 42 differences each. All were diffed with golden_diff against HEAD and reviewed, and all trace to the new cross-file call edges:

• fastapi direct downstream / Popularity Rank:
  • applications.py 44 → 191: test modules build FastAPI() at module level, and FastAPI lives there;
  • routing.py 13 → 30;
  • param_functions.py 92 → 103.
• What those move:
  • Composition Fit z-scores for 12 files, including a few non-Python files through the corpus-wide distribution;
  • within-repo file clusters (Repository Archetype and Ecosystem Fingerprint) in fastapi, numpy and wtfpython, via the PageRank / blast-radius features;
  • network_macro (modularity 0.849 → 0.824, avg path length 1.98 → 1.82);
  • three forensic top lists.
• No Structural Magnitude, coordinate or function-count changes.

Tests

• New tests/extraction/languages/test_python_module_level_unit.py:
  • module-level calls, qualifiers and receiver types;
  • unit bodies and class headers are not module calls;
  • a definitions-only file has no unit;
  • embedded_python gets it too;
  • the unit adds no impact.
• tests/tools_recorders/test_fcall_data.py: a synthetic unit persists to synthetic_unit_data (not function_data), rehydrates with calls_only, and resolves identically after rehydration.

Verification

$ pytest tests/                                               -- 11456 passed; the 1 failure (dead-key audit on a SQL column read) is fixed in 59ad9f25
$ PATH=.venv/bin:$PATH python tests/tools/audit_check.py      -- all clear (after 59ad9f25)
$ python tests/tools/call_graph_accuracy.py --ci              -- OK
$ python tests/tools/call_graph_resolution.py python --ci     -- OK
$ python tests/tools/tree_sitter_accuracy_audit.py --ci --all -- 30 languages, all OK
$ python tests/tools/crucible_check.py --mode both --update --yes   -- 42 differences each, reviewed above

Next in this line: decorator → decorated-function edges (@provide_bucket_name wraps every call to the method), then reference edges for functions passed as values (Depends(get_db), callbacks).

🤖 Generated with Claude Code

https://claude.ai/code/session_01GCpfMCUMpco98QGbsuv2sr

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab