--exclude-header cannot stop a transitively-included header being analyzed

Index: Known gaps — closed

Status (2026-10-11): CLOSED. Already fixed in code before this re-verification: the exclusion is honoured post-parse by declaring header, not only on the parse roots. extract/header_exclusions.py::scoping_header_predicate widens the dump's dependency scope (workflows/dump/dependency_scope.py) so a declaration whose source_header matches a pattern is scoped out like a toolchain header's, for both castxml and clang (each records the declaring file); workflows/header_exclusion_audit.py records a pattern as achieved when it matched only a transitively reached header. The headers are still parsed, so types the public header needs still resolve. Decision for a public declaration whose type lives in an excluded header: the type is kept and checked (direct-reference retention, same as a toolchain type a public signature names), because its layout is part of the public ABI. Disposition stays recorded (ADR-067): a coverage_warnings line names the excluded patterns, edge coverage carries dependency_headers_excluded, and an exported symbol declared only in an excluded header is reported as exported_not_public naming that header. No schema, default or exit code changed. Regression: tests/test_exclude_header_transitive_integration.py (public.h includes internal.h; internal-only layout/enum change reported without the flag, not with it; a public.h break still exits 4 with it; castxml and clang over g++-built libs), alongside tests/test_exclude_header_dependency_scope.py and the embedded-type case in tests/test_checkout_root_independence_integration.py. Out of scope: dump --include-system-declarations (full dependency scope) does not apply the transitive half, by that flag's own contract.

Update (2026-10): the descriptor half is moot. The ABICC compat front end, its XML descriptors and model/header_skip_rules.py were removed, so no run reads <skip_headers>/<skip_including> any more. The limitation below still holds for native --exclude-header, which shares it exactly.

The glob question this entry used to record is closed. It asked whether real ABICC globs in <skip_headers>, and left descriptor skips matched as exact basenames or paths in the meantime -- under which every tree-relative rule (Intel MKL's own fftw/fftw.h, fftw/offload/) matched nothing at all while the run still produced a confident verdict. The evidence the entry asked for is in ABICC's own source: Internals/Path.pm's classifyPath and Internals/Filter.pm's skipHeader_I classify each rule into one of three classes -- a bare name matched against the basename, a value containing a separator matched against the path at component boundaries (including a directory's descendants), and a metacharacter-bearing value compiled as a regex-like pattern. abicheck/model/header_skip_rules.py is that classification, and the descriptor's snapshot now records excluded_header_matching = "abicc" rather than "exact", since the same text genuinely names a different set of headers under the two rules.

Rewriting every rule to a bare basename was considered and rejected: it fixes MKL and over-excludes an unrelated version.h under a different subtree, and no example test drawn from the reported case can tell the two apart.

What is still open is the other half of <skip_headers>'s meaning. ABICC distinguishes two elements -- <skip_headers> is "do not include and do not analyze", <skip_including> is "do not include directly, but still analyze when reached" -- and the two are now modelled separately, with only <skip_headers> recorded as an achieved narrowing of the surface. Both correctly drop a header from the direct -H operand list. Neither can stop a header being parsed when another header reaches it through its own #include, because filtering happens on the resolved header list after the directory walk and before any parse. The native --exclude-header path shares that limitation exactly (extract/header_exclusions.py), which is why closing it is a post-parse filter by defining header rather than a change to either rule language -- and why it is recorded here rather than patched on the descriptor side alone.

Until then, docs/reference/abicc-format-compliance.md states partial support for <skip_headers>, and the two elements' remaining difference is in what a run records, not in what it parses.