Add JSONL conversion + clean preprocessing stages and uv support

[Zehui127]

Summary

• Two new Snakemake targets wrapping downstream preprocessing:
  • jsonl — parallel per-shard FASTA → json_sdiff JSONL conversion with a resumable per-shard cache, emitting {forwards,pairwise,combined}_{train,test}.jsonl per analysis.
  • clean — quality filter (max_hunk_len, gap_allele_frac, ref_gap_frac, mut_density) emitting filtered JSONL + per-file drop manifest CSV.
• scripts/fasta_to_jsonl.py (from pegasus-evals) and scripts/clean_lib.py (from pegasus-datasets) are vendored so the pipeline is self-contained — no sibling repos required.
• Both targets honour target_analyses and trajectory_mode. Defaults live under new jsonl: / clean: blocks in defaults/config.yaml; per-dataset overrides go under analysis.{name}.jsonl / analysis.{name}.clean.
• Adds pyproject.toml + uv.lock for uv sync / uv run snakemake .... requirements.txt is left intact for plain-pip users.

Test plan

• uv sync resolves and installs all deps
• uv run snakemake --configfile defaults/viral.yaml --cores 4 -p clean --config target_analyses='["n450-xs"]' runs end-to-end and produces 4 trajectory shards, 6 JSONL files, 6 cleaned JSONL files + manifests
• Dry-run -n shows the expected DAG for trajectory_mode=both (20 jobs) and reduced DAGs for forwards / pairwise only

🤖 Generated with Claude Code

[trvrb]

Thanks for the PR @Zehui127. Working through this with Claude Code, it flagged just one thing as significant

---

--context-size 16 actually produces 12-bp context (scripts/fasta_to_jsonl.py)

format_json_sdiff_hunk calls find_unique_context_size(min_context=context_size, max_context=12) with max_context=12 hardcoded. With CLI default context_size=16, range(16, 13) is empty — the loop never executes and it falls through to return 12. I verified at runtime: the hunks in forwards_train.jsonl have left-context strings of exactly 12 chars (e.g. 'TGGCAGGAATCT'), not 16.

The defaults/config.yaml value jsonl.context_size: 16 is silently a no-op above 12. Three possible fixes; the cleanest is probably making max_context track min_context:

  def find_unique_context_size(ref_seq, pos, ref_allele, min_context=8, max_context=16):
      for ctx_size in range(min_context, max(min_context, max_context) + 1):
          ...

…and updating the call site in format_json_sdiff_hunk to pass max_context=context_size (or a separate max_context_size arg). Worth deciding whether the CLI default should be 12 or 16 with the broader range — the original pegasus-evals semantics may inform that.

---

Does this make sense to you?

[trvrb]

Separately, I have a small pedantic issue: I've used to snakemake clean as a workflow endpoint that removes temporary build artifacts, in analogy to make clean. Can you rename this to something else? Maybe jsonl-qc?