Opaque-type suppression still falls back to bare RecordType.name when stable identities are incomplete

Index: Known gaps — closed

Status (2026-10-14): CLOSED. Both remaining mechanical parts are fixed. find_by_value_types's leaf widening now applies C++ unqualified lookup (compare/opaque_leaf_lookup.py): a bare leaf reference written inside a known scope does not expose ns::Handle when the snapshot declares a same-leaf type in a scope that lookup reaches first; with no such declaration, or no known scope, the reference stays exposed. find_opaque_struct_types' asymmetric-existence rule now pairs a forward declaration with a definition under either spelling (name or qualified_name), so a DWARF forward declaration keyed ns::Handle and a header definition keyed Handle/ns::Handle (or the reverse) are no longer read as "absent", and a visible declaration's qualified_name also blocks the spelling tier. Accepted by design, not a gap: a change with no identity whose spelling is shared only by opaque declarations is suppressed by spelling, since no identity or visible namesake contradicts it. Regression: tests/test_opaque_leaf_lookup.py.

Previous status (2026-10-13): PARTIAL. Fixed: a spelling that also names a visible declaration is never used to suppress through the spelling tier (OpaqueTypeIndex.shared_with_visible, both opaque-type and opaque-struct paths, complete or incomplete pairing); a stable id that a visible declaration also carries no longer proves opacity. Still open: (1) a change with no identity whose spelling is shared only by opaque declarations is still suppressed by spelling, by design, because no identity or visible namesake contradicts it; (2) the exposure-detection mirror, find_by_value_types leaf widening, which pins test_find_by_value_types_leaf_widening_bare_reference_in_another_scope_is_a_documented_gap in tests/test_cov95_diff_filtering.py; (3) a DWARF-only forward declaration paired with a full definition under the same name keeps the name-based criterion of find_opaque_struct_types, since DWARF resolves no identity to tell the two apart.

Previous status (re-verified 2026-10-09 at 456989f): PARTIAL. compare/opaque_types.py:74-130 now uses OpaqueTypeIndex carrying stable EntityIds plus a bare-name 'local' tier. contains(strict=True) narrows the bare-name collision only when intersect() proves the pairing is complete, and otherwise falls back to the old bare RecordType.name matching. diff_filtering.py:854 _root_type_name now prefers Change.qualified_name.

diff_filtering._find_opaque_types() (and its siblings _find_by_value_types()/_downgrade_opaque_type_changes()) index a snapshot's opaque/impl-private types by t.name alone, and _root_type_name() derives a Change's matching key from Change.symbol the same way diff_types.py stamps it — also the bare name for a top-level type change (name = t_old.name), never qualified_name. Two distinct records sharing a bare name in different namespaces (a complete, genuinely public api::Foo and an unrelated forward-only impl::Foo) therefore collide in the same opaque: set[str]: if impl::Foo is opaque, _downgrade_opaque_type_changes() silently suppresses a real TYPE_SIZE_CHANGED/field-change finding on the unrelated, complete, public api::Foo too, since both match the bare key "Foo". Confirmed by reading the code (no live repro run); this predates PR #719 and already applied identically to castxml's own is_opaque=True types — the PR's clang-backend opaque-stub fix (previously clang silently dropped every forward-decl-only type instead of emitting a stub) makes this reachable from a new source, not a new bug class. Not fixed here: a correct fix needs qualified identity threaded consistently through _find_opaque_types/_find_by_value_types/_downgrade_opaque_type_changes and _root_type_name's several call sites in diff_filtering.py (at least five, by grep), none of which currently have test coverage for the cross-namespace-collision case to validate a change against — a systematic, cross-cutting rework, not a scoped fix reactive to one review comment. Filed here rather than attempted under this PR's time budget, per this file's own "known gaps over risky reactive patches" convention.