bench: load committed Tier-A CMNN datasets from LFS (SOURCES lfs + runners use require_dataset)

[davorrunje]

Context

The CMNN paper-reproduction datasets (auto, heart, compas, blog, loan)
are now registered as Tier A in the honest-scholar registry datasets.yml
and their redistributable copies are committed via Git LFS under
benchmarks/data/<name>/{train,test}_<name>.csv.gz (author sign-off 2026-07-22;
see the PR that closed #138). honest-scholar dataset verify <id> resolves and
checks those committed bytes at their repo path.

However, the benchmark harness still loads these five datasets from the
runtime download cache, not from the committed copies. Every runner resolves
their data_dir via benchmarks.datasets.download.default_dest() (the
~/.cache/mononet/datasets / $MONONET_DATA_DIR cache populated by
benchmarks/datasets/download.py from Zenodo 7968969), and
benchmarks/datasets/sources.py::SOURCES has no entries for these five — so
resolve_dir("auto") / require_dataset("auto") would KeyError. The large-screen
datasets (adult, german, lc, polish, taiwan) already do the right thing:
they are in SOURCES with hosting="lfs" and load from benchmarks/data/ via
require_dataset. This issue makes the CMNN five consistent, so a fresh clone
reproduces CMNN results with zero network setup (git lfs pull only).

Goal

Make the committed Tier-A LFS copies the harness's canonical source for auto,
heart, compas, blog, loan, mirroring the existing large-screen datasets —
so the benchmark runners load them from benchmarks/data/<name>/ with no runtime
Zenodo fetch.

Where

• benchmarks/datasets/sources.py — SOURCES, resolve_dir, require_dataset.
  Add the five with hosting="lfs", license="CC-BY-4.0 (Zenodo)",
  url="https://zenodo.org/records/7968969", and a prep_hint pointing at
  benchmarks/datasets/download.py.
• The runners that currently pass data_dir=default_dest() for these datasets:
  • benchmarks/run.py:159
  • benchmarks/sensitivity.py:81
  • benchmarks/flavor_ablation.py:136
  • benchmarks/loan_size_ladder_run.py:244
  • benchmarks/_common/search.py:445
• benchmarks/large_screen_run.py:79 is the reference pattern (already uses
  require_dataset).
• Leave benchmarks/datasets/download.py + benchmarks/datasets/manifest.toml
  as the regeneration/provenance path (manifest.toml pins the raw upstream
  .csv SHA-256; datasets.yml pins the committed .csv.gz SHA-256).

Proposed approach

Add a single resolver that returns require_dataset(name) when name in SOURCES
(committed Tier A) and falls back to default_dest() otherwise (generator-backed
synth_* specs and any cache-only dataset), then route the runner call sites
through it. This keeps synthetic datasets (which ignore data_dir) and the
download path working. The committed .csv.gz decompresses to the exact Zenodo
raw .csv (verified when #138 landed), so loaded bundles are byte-identical to
the previous cache path — no result changes expected.

Acceptance criteria

• python -m benchmarks.run --dataset heart ... (and the other four) loads
  from benchmarks/data/<name>/ on a fresh clone with no prior download
  and no MONONET_DATA_DIR set (after git lfs pull).
• require_dataset("auto"|"heart"|"compas"|"blog"|"loan") resolves to
  benchmarks/data/<name>/ instead of raising KeyError.
• Loaded DatasetBundles are byte-identical to the pre-change cache path
  (spot-check X/y shapes + a checksum of the arrays for one dataset).
• uv run pytest -m "not slow" and the benchmark unit tests pass.
• honest-scholar dataset verify <id> still passes for all five.

References

• dataset: register the 5 CMNN paper-reproduction datasets with real SHA-256 + datasheets #138 (registration + LFS commit of the five datasets)
• datasets.yml, benchmarks/datasets/sources.py, benchmarks/datasets/download.py
• benchmarks/datasets/manifest.toml