Evidence-provider model — investigated, found not to reproduce as described; no fix applied

Index: Known gaps

Status (re-verified 2026-10-09 at 456989f): NEGATIVE_RESULT. Investigated and found not to reproduce. The modules it relies on still exist (dumper_layout_backfill.py, policy/classification.py:619 evidence_status_for_result).

A status-review follow-up asked whether evidence_status_for_result's report-level downgrade (kind-level ARTIFACT_PROVEN → UNATTRIBUTED only when DiffResult.evidence_tiers is header-only for the whole comparison) can let an individual header-derived BREAKING_KINDS finding read as artifact-proven merely because some other, unrelated part of the same report had binary evidence. Traced this for the highest-stakes family it could apply to — layout findings (TYPE_SIZE_CHANGED/TYPE_ALIGNMENT_CHANGED, diff_types.py) — and it does not hold up: (1) the direct-clang L2 backend's RecordType.size_bits/alignment_bits are populated only when dumper_layout_backfill.backfill_dwarf_layout() actually corroborates them against real DWARF (model.py's own dwarf_layout_coherence docstring) — with no DWARF to backfill against, those fields stay None and _append_type_size_and_alignment_changes's own is not None guard means no finding is even emitted, so an "unconfirmed clang-derived layout finding" cannot occur; (2) the castxml backend computes struct layout itself, via its own bundled real compiler targeting the resolved ABI — model.py already documents this as deliberately treated as sufficient L2 evidence ("trivially self-consistent by construction", not needing DWARF corroboration), a prior, intentional design decision this pass would have to overturn, not merely patch. The one place this class of risk is genuinely live is exactly the already-tracked toolchain-identity-probe gap above (castxml/clang invoked with compiler/ABI flags that don't match the real build) — not a separate evidence-status bug. A real per-finding provider model (recording, per Change, which of L0–L5 actually produced/corroborated it) would need new provenance plumbing through all ~45 Change(...) construction sites across diff_*.py/buildsource/*.py, each individually verified against the FP-rate/mutation-score gates — a multi-day project on its own, not attempted here.