docs(adr): nine retrospective architecture decision records (1/5)

[sqr00t]

Closes #11.

Adds nine ADRs under docs/adr/, plus an index — ten files in total. ADRs 0001–0008 are
retrospective, reconstructed from git history, issues and pull requests; 0009 records an open
question rather than a decision taken. No code changes — documentation only.

What is here

#	Title	Status	Date
0001	Wrap upstream bw's C++ rather than reimplement or shell out to R	accepted	2025-11-20
0002	Isolate Rcpp behind shim.hpp and keep the model sources unchanged	accepted	2025-11-20
0003	scikit-build-core + pybind11 + CMake as the build backend	accepted	2025-11-20
0004	uv as the packaging, environment and lock tool	accepted	2025-11-20
0005	Polars for the tabular output helper	accepted	2025-11-20
0006	Test strategy: unit/regression split, PHYS_RTOL = 1e-4, Brownian pinned by seed only	accepted	2026-07-09
0007	Multiplatform CI via tox-uv, and the GCC/libstdc++ compatibility fix	accepted (unmerged)	2026-07-09
0008	Publish to private AWS CodeArtifact via GitHub OIDC, and unset UV_INDEX locally	accepted	2026-07-09
0009	Evaluate a Rust reimplementation with PyO3 bindings	proposed	2026-08-18

The index in docs/adr/README.md groups these by theme, and the themes are ordered by date, so
reading top to bottom also reads chronologically. Numbering ascends with decision date, so the two
orderings coincide without renaming anything — the README states this as the rule for new ADRs.

What to check when reviewing

The Consequences sections are the load-bearing part, and they are meant to be uncomfortable. Please
push back if any of these overstates or understates the cost:

• 0001/0002 — byte-identical upstream sources mean we inherit upstream bugs and cannot reformat
  those files, so a repo-wide clang-format --style=Google can never pass over them
• 0006 — the Brownian path is pinned only by seed-reproducibility, because
  std::normal_distribution is not portable across standard libraries
• 0008 — the publish job builds a single-platform wheel, so consumers elsewhere compile from source

Every claim cites a commit sha or a file:line. If a citation does not resolve, that is a bug in the
ADR — please flag it.

ADR 0007 is marked accepted (unmerged) because the fix it documents is on #10, not yet on main.
The index's status legend explains that case.

One known follow-up: ADR 0007 describes the cpp format job as it stood in July 2026 —
non-blocking, with clang-format run repo-wide. #21 changes that job and #22 records the new policy
as ADR 0010, which corrects the description here. Reviewing this PR on its own terms is still right:
0007 documents the decision as it was taken.

Stack

This is the third of six PRs in the chain, and the first of the four docs PRs. It is based on
cpp/format-blocking (#21), not on main. Merge bottom-up.

main
 └── fix/gcc-cpp-compat            #10  compile fix + ColProxy constructor
      └── cpp/format-blocking      #21  .clang-format, pinned+scoped clang-format, job blocking
           └── docs/adr-records                    #15  nine ADRs  ← this PR
                └── docs/runbooks                  #16  four runbooks
                     └── docs/roadmap-and-rust-port #17  roadmap + Rust assessment
                          └── docs/entry-documents  #18  README/CONTRIBUTING/ONBOARDING
                               └── docs/cpp-format-policy #22  ADR 0010 + cpp-job corrections

The docs work is split into four PRs because the parts have genuinely different reviewers and
different failure modes: the ADRs need someone who was there, the runbooks need someone who will
follow them, the roadmap needs whoever sets priorities, and the entry documents need a reader who
does not know the repo.