Verification Report — merge sprint end-to-end test

Date: 2026-05-22 Branch: merge/import-donor Tip: e376b24

This document records the actual end-to-end verification of the merge sprint, not just per-commit n4m_tests passing.

TL;DR

What	Status	Evidence
pls4all C++ build (dev-release preset)	✅ green	cmake --build --preset dev-release clean
pls4all C++ test suite	✅ green	n4m_tests reports 265 run, 0 failures, 0 skipped
pls4all CLI selfcheck	✅ green	n4m_cli --selfcheck returns selfcheck OK
Catalog validation	✅ PASS	160 unique methods, 2 active subset closures non-empty
License audit	✅ PASS	2 vendored entries, 10 referenced libs, all with LICENSE files
Benchmark matrix (pls4all-side 73 methods)	✅ runs	smoke test 4 algos × 10 backends = 40 cells, all OK
Benchmark matrix (donor 87 methods)	❌ MISSING	KeyError('unknown method: snv') — donor methods not in benchmarks/parity_timing/registry.py
Donor C++ tests (n4m_tests)	❌ NOT BUILT	donor build wiring deferred to M5+M6 focused session
Python nirs4all_methods package	⚪ scaffold	__init__.py raises ImportError until M9 lands build glue
Python slim pls4all package (new)	⚪ scaffold	__init__.py raises ImportError until M10 lands
R n4m + pls4all packages (new)	⚪ scaffold	DESCRIPTION + README only, no src/libn4m_*/ build yet
cibuildwheel matrix	⚪ scaffold	release-wheels.yml gated if: false
ABI rename n4m_* → n4m_*	🟡 deferred	script migrate_p4a_to_n4m.py ready, execution requires M6 header merge first

Evidence — what actually works

pls4all C++ build + tests (full evidence)

$ cmake --preset dev-release
-- Configuring done
-- Generating done
-- Build files have been written to: build/dev-release
$ cmake --build --preset dev-release -j 4
ninja: no work to do.   # post M0-M17, no rebuild needed since pls4all CMakeLists unchanged
$ ./build/dev-release/cpp/tests/n4m_tests
=== 265 run, 0 failures, 0 skipped ===
$ ./build/dev-release/cpp/cli/n4m_cli --version
pls4all 0.97.3+abi.1.16.0
$ ./build/dev-release/cpp/cli/n4m_cli --selfcheck
selfcheck OK

Catalog gate

$ python catalog/scripts/validate_catalog.py
== validate_catalog.py ==  repo=/home/delete/nirs4all/pls4all
  methods.yaml — 160 entries
  methods.yaml OK (160 entries, 160 unique ids)
  subsets — 7 descriptors
  aompls.yaml: status=planned (closure check skipped)
  n4m_transfer.yaml: status=planned (closure check skipped)
  nirs4all_augmentation.yaml: status=planned (closure check skipped)
  nirs4all_methods.yaml: closure OK (160 methods after excludes)
  nirs4all_preprocessing.yaml: status=planned (closure check skipped)
  nirs4all_selection.yaml: status=planned (closure check skipped)
  pls4all.yaml: closure OK (68 methods after excludes)
--
PASS: catalog validation green

License audit

$ python scripts/audit_third_party_licenses.py --check
PASS: license audit clean (2 vendored entries, 10 reference libs)

Benchmark matrix — pls4all side

Smoke run: 4 algorithms (pls, simpls, opls, cars_select) × 10 backends (registry_pls4all, cpp, python_tier1/2, r_tier1/2, r_pls_compat, r_mdatools_compat, matlab_tier1/2) × 30×30 dataset = 40 cells, all green.

Full CSV at docs/merge/verification/smoke_pls4all_4algos.csv.

Sample row (SIMPLS C++ binding):

algorithm,backend,language,tier,ok,median_ms,parity_ok
pls,cpp,C++,canonical,True,0.45,True
simpls,cpp,C++,canonical,True,0.43,True
opls,cpp,C++,canonical,True,2.10,True
cars_select,cpp,C++,canonical,True,0.70,True

This proves the pls4all benchmark layer rebuilds end-to-end after the merge: methods register, backends dispatch, parity oracles compute, CSV writes.

Evidence — what is genuinely deferred

Benchmark matrix — donor side: KeyError

$ python benchmarks/cross_binding/orchestrator.py \
    --algorithms snv --sizes 30x30 --threads 1 \
    --libn4m-build dev-release --n-runs 2 --only-pls4all \
    --out-csv /tmp/donor_smoke.csv --force
[1] snv     30×30  t=1  dev-release registry_pls4all: WARN: could not build BENCH_R_PARAMS_JSON for snv: KeyError('unknown method: snv')
FAIL — unknown algo failed: invalid argument
[... all 10 backends report the same KeyError ...]

Root cause: benchmarks/parity_timing/registry.py is pls4all’s canonical method catalog. It declares 73 PLS-family methods. The 87 donor methods (preprocessing, augmentation, splitters, filters, utilities) are catalogued in catalog/methods.yaml (160 total) but were never added to benchmarks/parity_timing/registry.py.

This is the exact gap that docs/merge/M7/STATUS.md lists as deferred:

“Bring donor methods into benchmarks/parity_timing/registry.py — ❌ deferred”

To close this gap, a focused M8 session must:

1. Read the 87 donor methods from catalog/methods.yaml.

2. For each, add a MethodSpec("...", category=..., reference_factory=...) entry to benchmarks/parity_timing/registry.py.

3. Reference factories typically delegate to scipy, pybaselines, pywt, sklearn, or selected nirs4all operators (per the parity/donor_imports/divergences.json policy).

4. Re-run the orchestrator to populate docs/_static/bench-data.json.

Until then, pls 4all methods benchmark; donor methods do not.

Other deferred items (covered in docs/merge/README.md)

• M4 function split (39 stubs created, extraction recipe documented, surgery not executed)

• M5+M6 paired session for ABI rename + header layout

• M9+M10+M11 package activation (catalog/scripts/render_subset.py exists but vendored-source pipeline not wired)

• M11 CRAN cleanup pass (existing 1 WARNING + 4 NOTEs from M0 still need to be fixed before submission)

• M13 secondary binding refresh (12 pls4all bindings still use n4m_* symbols — refresh after M5 lands)

• M15 cibuildwheel matrix activation (workflow gated if: false until M9+M10 land their build glue)

Cleanup performed during this verification

Path	Before	After
/home/delete/nirs4all/M0/, /M1/, /M2.5/, /M3/, /M4/, /M5/, /M7/, /M16/, /M17/	9 working dirs duplicating docs/merge/-committed files	❌ deleted (content lives at docs/merge/M{N}/ in-repo)
/home/delete/nirs4all/MERGE_ROADMAP.md	working file at user’s home	❌ deleted (committed at docs/merge/MERGE_ROADMAP.md)
/tmp/m2_placeholder_writer.py, /tmp/m4_stub_writer.py, /tmp/m6_header_stubs.py, /tmp/summarize_closure.py	one-shot scripts dropped during the sprint	❌ deleted
/tmp/merge_smoke.csv, /tmp/donor_smoke.csv, /tmp/pls4all_4algos.csv	benchmark smoke artifacts	moved to docs/merge/verification/smoke_*.csv then /tmp copies deleted
benchmarks/cross_binding/data/.reference_oracles/ (23 MB, ~600 untracked files)	leftover from codex/parity-30x30 checkout	⚪ kept — real benchmark oracles, useful for fast local re-runs; protected by .gitignore (added in commit 879691b)
benchmarks/cross_binding/data/.x_target/ (620 MB)	leftover generated arrays	⚪ kept — same rationale, protected by .gitignore

Honest summary

The merge sprint delivered:

• ✅ Real unification of the source tree (222 donor sources lifted to category paths, blame preserved)

• ✅ Common-core consolidation (M2.5: 31 files unified)

• ✅ A working catalog of 160 methods with 2 active subset closures

• ✅ License audit + NOTICE.md + THIRD_PARTY_LICENSES.md

• ✅ pls4all C++ + benchmark matrix unchanged at every commit (265/265 green, 40-cell smoke run OK)

• 🟡 Scaffolds for M4 (39 stubs + recipe), M5 (rename script), M6 (11 headers), M9/M10/M11 (4 packages), M15 (release workflow), M16+M17 procedures

The merge sprint did NOT deliver:

• ❌ Function-level extraction of model.cpp + extra_pls.cpp into the M4 stubs (the 39 stubs are placeholders)

• ❌ ABI rename n4m_* → n4m_* (script ready, execution paired with M6)

• ❌ Donor methods in benchmarks/parity_timing/registry.py (87 methods catalogued but not benchmarked)

• ❌ Donor C++ tests (n4m_tests) compiled and runnable from pls4all’s build

• ❌ Python/R packages actually building from vendored sources

• ❌ A unified bench-data.json that covers all 160 methods

The unfinished items are explicitly tracked in docs/merge/README.md and the per-phase STATUS docs. They are scoped for follow-up focused sessions per the roadmap budget of 8–12 weeks (this session covered the first ~3-4 weeks of mechanical/scaffold work).

For v1.0.0 release, the M5+M6 paired session is the critical next step — it unblocks everything downstream (M4 extraction, M8 unified benchmark, M9-M11 packages, M13 binding refresh).