fix(reachability): anchor user-input entry-point patterns with word boundary

[gadievron]

fix(reachability): anchor user-input entry-point patterns with word boundary

Base: master · Type: bug fix · Finding: F9 (MED-HIGH)

What

utilities/agentic_enhancer/entry_point_detector.py — add a leading \b to two
USER_INPUT_PATTERNS entries:

• the FastAPI alternation (Query|Body|Form|File|Header|Cookie)\s*\( → \b(...)\s*\(
• ArgumentParser\s*\( → \bArgumentParser\s*\(

Why

The unanchored alternation matched any identifier ending in one of those words,
so ordinary library calls — res.setCookie(, PQsendQuery(, req.getHeader(,
MyArgumentParser( — matched as user-input sources and were seeded as false
remote-web entry points. Measured across the eval corpus: ~1,800 false seeds
(postgres 431, kubernetes 717, symfony 600, laravel 377). False entry points
inflate the reachable set and mis-label attack surface.

Qualified forms still match after the fix: fastapi.Query(, models.Header(,
argparse.ArgumentParser( — the . supplies the word boundary.

Tests

tests/test_entry_point_input_pattern_boundary.py (new): rejects
setCookie(/PQsendQuery(/getHeader(/parseMultipartFile(; still matches
standalone Cookie(/Query(/Body(/Header(.

• RED (pre-fix): setCookie(token) matched Cookie(.
• GREEN (post-fix): 2 passed; test_entry_point_detector.py + bindings 18 passed.

Reachability impact (verified)

Empirically, this only removes FALSE input entry points — genuine FastAPI
dependency seeds are retained. No legitimate reachability lost. (Library-collapse
behaviour F6 is untouched — this change is two regex lines.)

Upstream coordination

No open PR modifies this regex. PR #120/#75/#76 reference USER_INPUT_PATTERNS
only in added comments / new helpers; the regex line is unchanged on all of them.
Non-overlapping.

Author notes

• Fix line: the two \b-anchored patterns at the FastAPI and ArgumentParser
  entries.
• Input it now handles: calls like setCookie(/PQsendQuery( that previously
  produced false entry points.
• Likely pushback: "does \b drop fastapi.Query(?" — no; the . is a
  non-word char so the boundary holds (covered by a test assertion).

[gadievron]

Finding F9 (MED-HIGH) from a multi-language reachability audit. Standalone fix off master; no overlap with any open PR (the regex line it edits is untouched by #120/#75/#76). Removes ~1,800 false remote-web entry points (e.g. setCookie(, PQsendQuery() without dropping genuine FastAPI seeds — reachability-verified.