Retrospective: INT1 (integration of M1b and M4)

Status: complete (2026-09-28) on main. M1b (the CUDA sssp backend) was finished and accepted on main; M4 (the project infrastructure) was finished and accepted on the branch m4-infra, which ran in parallel. INT1 merges the two, makes the documentation site cover M1b, applies the author’s credit and citation facts, and records the merge policy in force. The branch m2-cycle (M2a, cycle_count on the CPU) is not part of INT1; it is merged in M2b.

What was done

Commit	What
92e5f54	git merge --no-ff m4-infra (“Merge branch m4-infra: M4 project infrastructure”). Three textual conflicts, resolved by combining both sides: CONTRIBUTING.md (M4’s guide, plus M1b’s CUDA presets and ci/gpu_local.sh; cuda-build joins the required checks), README.md (M1b’s CUDA build and parity sections, plus M4’s clone step for the originals, now for MOSP-CUDA too) and ci/check.sh (M4’s steps tidy, harness, DYNG_CHECK_ONLY and the Sphinx docs under M1b’s shared-lock heavy wrapper; --help prints the whole header)
37f7a5d	cuda-build.yml (from M1b) meets M4’s workflow checks: timeout-minutes: 120, persist-credentials: false, one grouped redirect to GITHUB_ENV (actionlint’s shellcheck SC2129)
b450a45	examples/cpp/first_update.cpp (from M4) uses graph<> (the int32 default of ADR 0009, decided in M1b), reads the result with to_vector() and runs on cuda; CTest example.first_update.cuda (label gpu, skipped without a device)
d7801ba	the site covers M1b: an API page for the new Doxygen group generators; the core page names the CUDA pieces; install, choose-a-backend, backends-and-resources, getting-started, roadmap, tutorial and the short plan describe the CUDA backend instead of “arrives in M1b”; the algorithm tables (README, landing page, algorithms index) agree; the parity how-to has the CUDA replay and links to the M1a and M1b certificates
cace0ff	the author’s facts of 2026-09-27: ESCHER (IPDPS 2026) title and authors, TruCy as a submitted manuscript, S M Ferdous’s affiliation without an e-mail address, the placeholder-identity commits credited to S M Shovan, no funding line; GOVERNANCE.md approvals log
a44bc71	the merge policy (ADR 0019) in GOVERNANCE.md, CONTRIBUTING.md, the pull request template and the repository settings guide, which also marks the settings applied on 2026-09-27 and lists the real job names of cuda-build and docs among the required checks
59fc861	CHANGELOG

Things the merge brought together that needed no edit: M1b’s ADRs (0003, 0009, 0015-0018) and retrospective reach the site through the globs of docs/adr/index.md and docs/developer/retrospectives/index.md, and their rows were already in docs/adr/README.md; M1b’s CUDA headers (stream.hpp, buffer.hpp, copy.hpp, memory.hpp) are in the core group, so the Breathe page docs/api/cpp/core.md renders them (stream_ref, cuda_async_memory_resource, pinned_host_memory_resource, to_vector); the sssp page of M1b builds under -W -n with one exception (below); parity/README.md merged without a conflict and is included by docs/developer/parity.md.

Deviations (each a pragmatic choice with the plan’s intent)

Plan	What was done	Why
Sections 8.9 and 10.1: squash merges only	ADR 0019: merge commits for milestone and integration pull requests, squash for external contributions, no rebase merging; the main ruleset allows Merge and Squash and does not require a linear history	the parity certificates, retrospectives and ADRs cite milestone-branch SHAs (for example the M1b gate records and M4’s trial merges); a squash would leave them outside main. This is the policy the lead maintainer set in the repository settings on 2026-09-27
M4 notes: the sssp page links ../../parity/results/M1b.md (M1b wrote it before the site existed)	the link goes through the repository URL (github.com/dyng-dev/dyng/blob/main/parity/results/M1b.md), which ci/docs_links.py checks	files outside docs/ are not pages of the site; M4’s convention for them is the repository URL
Section 3.2: CITATION.cff references per paper	TruCy is type: unpublished with status: submitted and a note naming the journal (BibTeX @unpublished)	the author asked for a manuscript, not an article; CFF 1.2.0 has no manuscript type (cffconvert --validate rejects it), and unpublished is the schema’s type for a manuscript
(no plan item) the parity harness tests in ci/check.sh	pytest of the harness step runs under the shared lock like the other heavy steps	the machine rules: test suites take the shared lock

Validation

Fresh clone of main at 59fc861 (the commit before this retrospective), GPU 1 for the tests, the shared lock for every heavy step:

Check	Result
ci/check.sh --parity	all checks passed in 2 minutes: clang-format; cpu-only 234/234 and dev 246/246 (ctest -L cpu); clang-tidy naming; REUSE; provenance (88 files); harness 26 passed; docs (Doxygen, coverage of 105 compounds, Sphinx -W -n, docs_links over 1156 files); pre-commit (actionlint, zizmor, the .github metadata check on cuda-build.yml included); parity preset golden replay
ci/gpu_local.sh (dev-cuda)	all steps passed in 8 minutes: build; ctest -L gpu 92/92 (with example.first_update.cuda); ctest -L cpu 246/246 in the CUDA build; the sssp golden corpus on cuda, 495/495 cases byte-equal; compute-sanitizer memcheck (0 errors, 0 leaks) and synccheck (53/53, 0 errors); clang-tidy on the CUDA branches
ci/docs.sh (inside ci/check.sh, and again on the final commit with this page)	builds the site with every M1b page: ADRs 0003, 0009, 0015-0019, the M1b retrospective (and this one), the CUDA headers on the core page (stream_ref, cuda_async_memory_resource, pinned_host_memory_resource, buffer, to_vector) and generators::legacy on its own page
cffconvert --validate -i CITATION.cff (cffconvert in a scratch venv)	valid against CFF 1.2.0
git grep -nE '^(<<<<<<<|=======|>>>>>>>)( |$)'	empty; git branch --merged main lists m4-infra
read-only GitHub API (gh api repos/dyng-dev/dyng, .../actions/permissions*, .../rulesets, .../environments, .../private-vulnerability-reporting, .../vulnerability-alerts, .../automated-security-fixes, gh api orgs/dyng-dev, orgs/dyng-dev/installations)	the settings marked done in the guide’s checklist; no ruleset or branch protection on main; no setting was changed

Open items and notes for the next steps

• The main ruleset (repository settings guide, step 9) waits for the first pull request: every required check (including the three cuda-build jobs, site and DCO) must have run once. The next milestone merge (M2b, the branch m2-cycle) is the natural first pull request; it is merged with a merge commit (ADR 0019).

• Hosted runs. cuda-build.yml and docs.yml have not run on GitHub yet (only lint and cpu have, on the M1a history); the first push of main runs them. The labels of .github/labels.yml are applied by the labels workflow on that push. After it, run the docs workflow once with Run workflow for the external-link job (M4 open item).

• Settings still open (guide): the optional restriction of repository creation in the organization (step 1.3), CodeQL (optional, step 6.5), the Discussions categories and the pinned roadmap (step 12), Read the Docs and Zenodo (checkpoint A4).

• ADR numbers. The next free number is 0020; m2-cycle must not reuse 0019 when it is merged in M2b.

• The algorithm table still exists three times (README, landing page, algorithms index) until scripts/regen.py generates it (M3); INT1 edited all three by hand.

• cycle_count is described as “planned; the CPU backends in progress on a branch” until M2b merges m2-cycle; M2b updates the three tables, the roadmap and the short plan.

• Author decisions still open (GOVERNANCE.md): O3 (the NOTICE institution line and years) and whether a co-author wrote code of the originals outside git. The funding acknowledgement is added by the author later.