feat(report): stamp reports with a reproducibility/provenance block (closes #31)

[dchaudhari7177]

Closes #31.

A card stated effect and cliff numbers but not the conditions they were measured under. Provenance now records n_seeds, model_id, repeng_version, hardware and wall_clock_s, renders at the foot of the markdown and HTML cards, and is exposed through --json.

Rendered block

## Reproducibility

| field | value |
|---|---|
| seeds per point (min) | 2 (contract: ≥ 3) |
| source model | org/demo-7b |
| repeng version | unknown |
| hardware | A100-40GB |
| sweep wall-clock | 612.5 s |

⚠️ **Measured over 2 seed(s): below the ≥ 3-seed reproducibility contract, so the
± spreads here are not a reliable error estimate.**

And in --json (paths stay flat, so existing consumers are unaffected):

"provenance": {"n_seeds": 2, "model_id": "org/demo-7b", "repeng_version": null,
               "hardware": "A100-40GB", "wall_clock_s": 612.5,
               "min_seeds": 3, "below_seed_contract": true}

Design decisions worth reviewing

• Frozen dataclass, not pydantic. The issue suggested a pydantic model, but core declares dependencies = [] ("core stays light"), and report.py models every other type as a frozen dataclass — pydantic here would be the first core dependency and would break test_no_optional_deps_imported's spirit. Say the word and I'll switch it.
• n_seeds is the minimum across dose and layer points, not the mean: the weakest point bounds the claim.
• The seed count cannot be spoofed. provenance_from() always recomputes n_seeds from the parsed curves and only passes the caller's other fields through, so what is printed on an artifact can't disagree with the CSVs it came from. A test asserts a caller-supplied n_seeds=999 is overridden.
• Unknown ≠ violation. Missing fields render unknown rather than failing, and an unknown seed count is deliberately not a contract breach — only a known count below MIN_SEEDS warns.
• No breaking change. build_report() keeps its signature and return type; provenance is an optional keyword.

Incidental fix

--model was documented as "recorded in output; informational" but was never actually read anywhere. It now feeds model_id as the fallback when no vector is passed, with a loaded vector's own metadata taking precedence.

Also adds --hardware / --wall-clock-s for the GPU sweep to supply, and a stderr warning when a card is built below the seed contract.

Tests

8 new CPU-only tests, no model download: min-seed selection, pass-through vs. recomputation, unknown defaults, the below_seed_contract boundary (parametrised over 1/2/3/4/None), markdown + HTML rendering, the low-seed warning appearing in both, as_dict(), and that summarise_provenance() agrees with what the card renders.

pytest -q                 436 passed
ruff check .              All checks passed!
ruff format --check .     20 files already formatted
mypy src                  Success: no issues found in 5 source files

One note on mypy: with the repo's configured python_version = "3.11" it aborts early on this machine with numpy/__init__.pyi:737: Type statement is only supported in Python 3.12 and greater — reproducible on a pristine tree, so it is a local numpy-stub artefact, not from this change. The Success above is from mypy src --python-version 3.12, which gets the stub to parse so the actual source is checked.

🤖 Generated with Claude Code