A failed header parse on PE or Mach-O falls back to export-table mode and can exit 0¶

Status (recorded 2026-11-01): OPEN. False-clean risk.

Evidence¶

workflows/dump/header_scoped.py (_try_header_scoped_dump, the direct PE/Mach-O route) wraps the header-scoped dump in except Exception, emits a UserWarning ("Header-based ABI scoping unavailable ... Falling back to export-table mode -- --header/--include were ignored") and returns (None, "header-backend-unavailable"). The caller then dumps the export table only. The comparison records requested_depth: binary (requested_depth_source: implicit) and header_context_status: not_evaluated, and a header-only change is simply not seen, so the run can exit 0. On ELF the same missing-include failure fails the run.

Reproduced on the macOS integration lane: in tests/test_sided_include_labelled_composition.py, the negative control that drops the shared include roots exits 0 on macOS and fails on Linux. That test pins the darwin behaviour (the warning, and binary evidence depth) until this is fixed.

Intended resolution¶

An explicitly supplied --header whose parse fails is a failed acquisition (AGENTS.md "Weaker evidence narrows conclusions": a failed extractor is an error or an explicit FAILED fact, never an empty surface), on every binary format. The narrower fallbacks that are deliberate (the MSVC mangling mismatch, mangling-fallback) need their own ruling, because a header set that parses but matches no export is a different state from one that does not parse. Flip the darwin branch of that test's negative control when this is fixed.