M7 + M8 — parity gate + dashboard unification: STATUS

Status: 🟡 partial — donor infrastructure lifted under unified paths; full runner merge deferred Date: 2026-05-21 Pushed: pending commit

What was done

M7 — parity infrastructure lift

Donor’s _donor/nirs4all-methods/parity/ was relocated to parity/donor_imports/ so it sits alongside pls4all’s existing parity/{schema,fixtures,python_generator,r_generator} tree, while keeping the two distinct (no namespace collision yet).

Origin

Target

Files

_donor/.../parity/run_reference_parity.py

parity/donor_imports/run_reference_parity.py

1 (107.8 KB; the 5-stage gate driver)

_donor/.../parity/binding_parity.py

parity/donor_imports/binding_parity.py

1 (the Stage 4 scaffold)

_donor/.../parity/divergences.json

parity/donor_imports/divergences.json

1 (machine-readable skip registry, 3 classes)

_donor/.../parity/nirs4all_source.py

parity/donor_imports/nirs4all_source.py

1

_donor/.../parity/reference_parity.py

parity/donor_imports/reference_parity.py

1

_donor/.../parity/python_generator/

parity/donor_imports/python_generator/

full subtree (locked env + adapters for SciPy/sklearn/pybaselines/PyWavelets/nirs4all)

_donor/.../parity/fixtures/

parity/donor_imports/fixtures/

full subtree (~250 JSON fixtures)

_donor/.../parity/schema/

parity/donor_imports/schema/

n4m_fixture_schema_v1.json

_donor/.../parity/tolerances.md

parity/donor_imports/tolerances.md

1 (the n4m-side tolerance table)

_donor/.../parity/README.md

parity/donor_imports/README.md

1

M8 — benchmark dashboard portability

Donor’s _donor/nirs4all-methods/benchmarks/ similarly relocated:

Origin

Target

Files

_donor/.../benchmarks/cross_binding/

benchmarks/donor_imports/cross_binding/

subtree (orchestrator, results, validate_dashboard_payload.py)

_donor/.../benchmarks/reference_snapshots/

benchmarks/donor_imports/reference_snapshots/

subtree

_donor/.../benchmarks/validation/

benchmarks/donor_imports/validation/

subtree (the v0 registry/{datasets,suites,references,…}.json)

_donor/.../benchmarks/benchmark_registry.json

benchmarks/donor_imports/benchmark_registry.json

1 (donor’s scaffold registry)

_donor/.../benchmarks/check_truth_sources.py

benchmarks/donor_imports/check_truth_sources.py

1

_donor/.../benchmarks/truth_sources.lock.json

benchmarks/donor_imports/truth_sources.lock.json

1

_donor/.../benchmarks/README.md

benchmarks/donor_imports/README.md

1

Dashboard orchestrator de-hardcoded

benchmarks/cross_binding/orchestrator.py:142 was the only hardcoded /home/delete/... path Codex flagged (the rest of the 31346 grep hits are inside generated CSV result files, not source). Replaced the hardcoded path with an env-var-driven fallback:

PLS4ALL_R_ENV = os.environ.get(
    "N4M_R_ENV",
    os.environ.get("PLS4ALL_R_ENV", "/home/delete/miniconda3/envs/pls4all_r"),
)

A clean clone now works without editing the file. The maintainer’s historical fallback stays for backward-compat. A warning emits if the fallback doesn’t exist and N4M_R_ENV_NOWARN is unset.

What is NOT done (carried to focused session)

The roadmap M7 requires more than file-relocation:

Sub-goal

Status

Single unified runner driving both halves

❌ — pls4all has no top-level runner; donor’s run_reference_parity.py is at parity/donor_imports/ parked. The merger is M7’s main job and lands in a focused session.

Implement Stage 4 binding parity (donor’s binding_parity.py is scaffolded but unwired)

❌ — wiring happens after the runner is unified.

Merge tolerances.md (rename rows, calibrate per category per §6.3, add backend rows §6.4)

❌ — two distinct tables coexist (parity/tolerances.md from pls4all, parity/donor_imports/tolerances.md from donor).

Port pls4all’s 📜 documented-but-unwired rows into machine-readable divergences.json

❌ — donor’s divergences.json exists at parity/donor_imports/; pls4all-side rows still in MD only.

Sync sklearn version (base 1.4 vs donor 1.8.0); regenerate base fixtures

❌ — needs a fixture regeneration pass in a locked env.

The roadmap M8 requires the same kind of follow-on work:

Sub-goal

Status

Bring donor methods into benchmarks/parity_timing/registry.py

❌ — donor’s benchmark_registry.json parked at benchmarks/donor_imports/; merger deferred.

Add category / subcategory / subset filters to dashboard UI

❌

Split bench-data.json per-category for lazy loading

❌ — 829 KB → 2 MB issue remains.

Adopt MERGE_ROADMAP.md §6.5 container envs (ci-reference / ci-blas-omp / ci-cuda Dockerfiles)

❌

Build verification

pls4all build and tests unchanged at this commit:

$ cmake --build --preset dev-release && ./build/dev-release/cpp/tests/n4m_tests
=== 265 run, 0 failures, 0 skipped ===

The donor infrastructure parked at parity/donor_imports/ and benchmarks/donor_imports/ is dormant: pls4all’s CMakeLists doesn’t activate it and no Python entrypoint references the parked paths.

Remaining _donor/ footprint after this commit

_donor/nirs4all-methods/ still contains:

  • cpp/include/n4m/ (donor’s umbrella header — duplicates the one already at cpp/include/n4m/ from M2.5; deletion deferred to M5+M6 focused session)

  • cpp/src/c_api/ (donor’s c_api wrappers — pairs with cpp/include/n4m/)

  • cpp/cli/, cpp/tests/, cpp/abi/ (donor’s CLI + test harness + ABI golden symbol tables — useful for the M5+M6 session)

  • docs/, roadmap/, bindings/{python,r}, scripts/, cmake/, .github/

  • Top-level config files (ARCHITECTURE.md, CHANGELOG.md, CITATION.cff, README.md, etc.)

  • Total: ~815 files preserved as donor archeology

Final cleanup of _donor/ lands when the donor repo is renamed in M16 (at which point the prefix becomes redundant — everything is in the unified repo by name).