Skip to content

G2 — Build-config matrix into compare, and bundle completion

Registry: UC-WF-probe-matrix (complete), UC-WF-bundle (complete), UC-TC-cxx-standard-floor (complete) Effort: M · Risk: medium (verdict-composition semantics)

Problem

This plan is now complete. It originally tracked two capabilities that existed but were not reachable from the mainline gate:

  1. Build-config matrix — abicheck/probe_harness.py + diff_build_config.py detect API_DEPENDS_ON_CONSUMER_ENV, CXX_STANDARD_FLOOR_RAISED, and BEHAVIOURAL_DEFAULT_CHANGED, but only via the separate abicheck probe command. They now feed compare/compare-release via --probe-matrix-old/new.
  2. Bundle analysis — abicheck/bundle.py detects cross-DSO breakage, but compare-release wiring is now complete for explicit cohorts, including case84.

Goal & acceptance criteria

  • [x] compare/compare-release merge matrix findings into the verdict (worst-of), with the matrix ChangeKinds appearing in the report. Shipped as --probe-matrix-old / --probe-matrix-new (pre-built matrix snapshots from abicheck probe run) rather than an inline --probe-spec: running a matrix needs compilers, so it stays a separate probe run step that feeds the comparison, keeping the compare commands hermetic. On compare the findings join the change list (JSON + SARIF); on compare-release they are release-global, so they run through the same checker.compare pipeline (over empty snapshots) — --suppress rules and --policy-file overrides apply identically to both commands — then fold into the worst-of release verdict and surface as a matrix_findings section in the JSON/markdown summary and as a dedicated testsuite in JUnit output. Verified end-to-end for both commands in tests/test_probe_examples.py and at the unit level in tests/test_cli_split_modules.py.
  • [x] Case 98 (CXX_STANDARD_FLOOR_RAISED) reaches its intended verdict through the mainline command (JSON + SARIF), not only probe compare. Case 97 (API_DEPENDS_ON_CONSUMER_ENV) now also fires end-to-end: the harness gap is closed — parse_elf_metadata falls back to .symtab when a relocatable probe .o has no .dynsym, so the object's defined global symbols are captured and the detector fires over the real compiled surface, reaching the mainline compare output (tests/test_probe_examples.py, tests/test_elf_object_surface.py).
  • [x] compare-release emits bundle_soname_skew; case84 lost skip: true and is validated end-to-end (tests/test_bundle.py::TestCompareReleaseBundleE2E). The check is opt-in via --bundle-cohort PREFIX (repeatable): cohorts are declared, never inferred from filenames, so an ordinary release that bumps one independent library while a sibling lags is not a false positive.
  • [x] Two additional self-contained probe specs under examples/probes/ (feature_macro.yaml, cxx_standard.yaml) with an end-to-end test (stock cc/c++, no external toolchain).

Design

  1. Matrix-into-compare: add a --probe-spec option to compare_cmd (abicheck/cli.py) and compare-release. When present, run run_probe_matrix() for each side, diff_matrix() the pair, and append the resulting Changes to the DiffResult before compute_verdict. Verdict composition is already worst-of, so no policy change is required; matrix kinds are already classified in change_registry.py.
  2. Confidence: when probes are partial, set DiffResult.confidence=low and add a coverage_warning (mirror probe compare --allow-failures).
  3. Bundle wiring: finish the compare-release → bundle.py path so detect_bundle_soname_skew() and the other bundle detectors run on the per-library cohort; surface bundle findings in the summary report. Remove skip: true from case84 and add the gen_bundle.sh build to CI.

Files & surfaces

  • abicheck/cli.py, abicheck/cli_compare_release.py (--probe-spec, bundle wiring).
  • abicheck/service.py (run_compare accepts an optional matrix).
  • abicheck/bundle.py (cohort entry point from compare-release).
  • examples/probes/*.yaml (new specs); examples/case84_bundle_soname_skew/.

Tests

  • Unit: matrix merge into DiffResult; verdict worst-of with a matrix kind.
  • @pytest.mark.integration: probe build + compare --probe-spec; bundle skew via gen_bundle.sh.
  • Update ground_truth.json for case84/97/98 and the autodiscovery harness.

Out of scope

Non-Linux bundle analysis (no DT_NEEDED/.gnu.version_* equivalent — track under G1). Auto-deriving a matrix without a spec.