feat(swing-sim): land variation execution-metadata orphaned by the slice pivot

[dieterolson]

Part of #4142

Why these files had no route in

The 34-PR variation consolidation (#4447) is being superseded piecemeal — main is absorbing that campaign as slices, so #4447 will likely never merge. That left this work stranded:

• it exists on consolidated/variation-morris-2026-08-13,
• its two earlier codex PRs (feat(variation): bind resolved execution metadata #4428, feat(variation): bind runtime replay identity #4431) are both closed unmerged,
• and the sibling consolidation Rate of Closure remainder: club builder, impact tensor, flight, multi-view workspace, ground and web companion (consolidates 43 PRs) #4466 does not carry it.

What this adds

The execution-metadata sidecar that pins what a variation plan actually ran against — resolved variable values with unit and dimension semantics, RNG algorithm and stream-derivation identity, and solver/runtime implementation identity — so a historical plan can be replayed with its reproducibility either proven or explicitly refused. Legacy plans without a sidecar resolve against the current registry behind an explicit warning rather than silently claiming reproducibility.

File	LOC
variation/_execution_metadata_schema.py	75	new
variation/execution_metadata.py	477	new
variation/tests/test_execution_metadata.py	300	new

Why the scope is 9 files and not 3

The premise that this slice's dependencies were already on main was package-scoped and wrong in three ways. Each was verified against origin/main rather than assumed:

1. variation/spec.py has no MAX_SAFE_INTEGER. execution_metadata raised ImportError at its import line — the module was not merely untested on main, it could not be loaded at all.
2. variation/registry.py has no VariableDef.dimension, which failed 20 of 26 tests with AttributeError.
3. The four variation_execution_document_*.json fixtures live outside the package, under src/rate_of_closure/web/src/model/__fixtures__/ — so counting the 40 variation/** files could not see them.

So spec.py (+40/−10) and registry.py (+30) come along, adding the safe-integer bound on n_runs/seed, signed-zero JSON normalisation, and the registry dimension field. This is the minimum coherent set: it was determined empirically, not guessed — adding these is exactly what turns the suite from 20 failures into 26 passes.

Dead __all__ removed

execution_metadata.py carried two __all__ assignments (an artifact of splitting the schema module out) and the second silently won, leaving the first eight-name block unreachable. Those eight names are a strict subset of the survivor, so no export changes; all 21 remaining names were verified to resolve.

Deliberately excluded

scripts/benchmark_rate_ensemble_archive.py is not in this PR. It imports ensemble_archive, ensemble_request_identity and ensemble_trace_authority — three of fifteen rate_of_closure/variation modules absent from main — and belongs to a separate ensemble-archive slice. Landing it here would add a script that cannot import.

Gates

$ python -m ruff check .
All checks passed!

$ python3 scripts/check_file_size_budget.py --max-loc 500 \
    --baseline-file scripts/monolith_baseline.txt --changed-only --base-ref origin/main
Scanning 5 changed files vs origin/main
File-size budget passed: 0 violations (max 500 LOC).

$ python -m pytest src/shared/python/swing_sim/variation/tests/test_execution_metadata.py -q --timeout=180
26 passed

No regressions from touching the two shared modules: 265 variation tests and 1183 swing_sim tests pass.

One inherited red: ruff format --check

python -m ruff format --check . reports 4 files — all src/data_processing/..., none touched here. This fails on main today, so every open PR inherits it. Addressed separately in #4525, which should merge first; this branch then goes green. My own files are clean:

$ python -m ruff format --check <the 5 python files in this PR>
5 files already formatted

tests/rate_of_closure/ also has ~8 pre-existing flaky failures (Qt/perf tests under the -n auto xdist config). Confirmed unrelated: the same failures occur on a clean main baseline, and the failing set differs run-to-run.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.