dSYM artifact reaches no dump path¶

Index: Known gaps

Status (re-verified 2026-10-25): OPEN. abicheck/dumper.py:1051 _dump_pe has no debug_roots parameter, and dwp_path/dsym_path are set only by debug_resolver.py and read only by extract/debug_artifact.py's description (no dump consumer). The directory form still resolves PDB/DWP/dSYM artifacts that are then dropped.

debug_resolver's chain can resolve four kinds of artifact, and only one of them is consumed. service_dump_native._dump_elf reads DebugArtifact.dwarf_path and nothing else -- a dwp_path, a dwo_dir or a dsym_path is resolved, logged, and dropped -- and the PE branch of _run_dump_uncached never passes debug_roots to _dump_pe at all, so a PDB found in a debug root is dropped too (the PE path's only PDB input is debug.pdb_path, via locate_pdb(pdb_path_override=...)).

Pre-existing, and predates plan Phase 7n -- --debug-root had the same shape. What 7n changed is that the gap became sayable: merging the debug role into one input meant naming a detached artifact directly, which would have resolved a PDB/DWP and then silently ignored it, and (first review round) made the resolver return one ahead of EmbeddedDwarfResolver, so a binary carrying perfectly good DWARF would have been compared with none. Both are closed: extract/detached_debug.DetachedDebugFileResolver is DWARF-only and yields to the rest of the chain, and naming a PDB/DWP file is a usage error naming debug.pdb_path (the spelling that is read).

The directory form is deliberately still accepted, because rejecting it would break the case that works -- a build-id tree or path mirror of ELF .debug sidecars, which is what the input is mostly for. So a directory that happens to hold PDBs still resolves to an artifact nobody reads. That is this gap, unchanged; do not "fix" it by rejecting directories.

Closing it properly means threading debug_roots into _dump_pe and teaching _dump_elf to consume dwp_path/dwo_dir/dsym_path -- real capability work on the extraction paths, not a CLI change, which is why ADR-068's flag-topology slices leave it alone. Found by Codex review on PR #1253.

Update (plan Phase 7n): --debug-root, --devel-pkg and --probe-matrix no longer exist as flags -- each merged into the one input for its evidence role (--debug-info, -H/--header, --build-info). The rejections above are unchanged in substance and still keyed on the generated destinations (devel_pkg2, probe_matrix_*, ...), which is what made them survive the rename; only the spelling each message names moved. Writing the test found one the manual audit had missed (the compare flag that is now the consumers.manifests config key), which is the argument for the mechanism in one line.

The whole compute/render split is new in this pass too: report/no_baseline.py now follows abicheck/report/AGENTS.md's convention — one compute_no_baseline_document resolving per-finding verdict/category, evolution counts, coverage and exit contributions, and pure render_* halves that format and decide nothing — so a new audit report section goes in the document, never into one renderer. The audit publishes its own schema (abicheck/schemas/audit_report.schema.json, audit_report_schema_version starting at 1.0) rather than a version of the compare report's: findings[], cross_source_evolution, exit_axes and a candidate-only pattern_preprocessor_scan block are its own, and changes: [] stays, so a consumer reading changes off any abicheck report still finds it. It briefly stamped report_schema_version instead, which offered the audit under the compare report's identity -- see that field's own note in report/no_baseline_document.py for why that is not merely cosmetic.