Cluster-ablation splicer for leave-one-cluster-out

[cmungall]

Cluster-ablation splicer for leave-one-cluster-out

Implements the splicer designed in #4. Given a hold-out set of datasets, it
produces an executable copy of the expert monolith
(data/gold/expert_code/harmonize_ess-dive_soilmoisture_data.py) with those
datasets' blocks removed, so the held-out cluster is absent from the agent's
exemplar context and any regenerated exemplar outputs.

How it works

The monolith is a series of # %% cells; the 19 per-dataset blocks are each
exactly one cell anchored by a single idx = N. The splice drops whole cells
whose idx is held out and keeps everything else verbatim (config, shared
helpers, the REF_IDX = 0 reference reads, and the location-dedup footer), so
the result stays runnable.

• Input: --holdout accepts integer indices or dataset_identifiers
  (resolved via the gold mapping JSON). No coupling to cv_folds.yaml.
• Safety by construction: requiring each hold-out to correspond to a real
  block automatically rejects REF_IDX (0) and the excluded datasets
  (11–14, 19–22), which carry no block — a clear ValueError otherwise.
• Output: the ablated .py to --output (or stdout).

Example

python -m src.folds.ablate_monolith --holdout 1,2,3,6,16,27 -o ablated_c1.py
# wrote ablated_c1.py (removed blocks [1, 2, 3, 6, 16, 27]; 13 blocks remain)

Tests

• 14 pytest cases + 2 doctests, all green. A synthetic mini-monolith fixture
  keeps the suite independent of the full corpus; additional cases run against
  the real gold file when present (19 blocks; cluster ablation stays
  ast.parse-able with REF_IDX/footer preserved; rejects 0/11/19).

Deferred

Per the discussion in #4, verifying the ablated script runs end-to-end is
gated on staging the raw inputs (separate work). This PR validates the splice
(correct blocks removed, output parses, scaffold preserved); the run-to-verify
step follows once ~/ess-dive_wfsfa_soil_datasets/ is populated.

Refs #4

🤖 Generated with Claude Code