feat: add an opt-in MyGene.info backend for the Ensembl pull

[SkyeAv]

Adds an opt-in MyGene.info backend for the Ensembl pull. The existing BioMart harvest
(src/datahandlers/ensembl.py) queries every Ensembl species dataset (200+), which is slow,
flaky (config already skips a "stuck" dataset), and impractical on a machine that cannot run a long
network harvest. The new backend pulls only the well-covered model organisms from the MyGene.info
BioThings API — a registered Translator Knowledge Provider — via a scrolling query, with no bulk
download.

MyGene backend

• New module: src/datahandlers/ensembl_mygene.py (pull_ensembl_via_mygene) pages MyGene with
  fetch_all/scroll_id and writes the same ENSEMBL/<dataset>/BioMart.tsv artifacts (BioMart
  display-header columns), so write_ensembl_gene_ids / write_ensembl_protein_ids work unchanged.
• Opt-in wiring: selected by config.yaml ensembl_source (default biomart, unchanged) and
  ensembl_mygene_taxa; dispatched in the get_ensembl rule (src/snakefiles/datacollect.snakefile).
  Production behavior is untouched unless the default is flipped.
• Coverage: human/mouse/rat/zebrafish/fly/worm/xenopus (verified live). Yeast (~2 genes) and dicty
  (0) are near-empty in MyGene and stay on BioMart; MyGene has no SGD field.

Correctness traps handled

• str-or-list normalization: MyGene returns ensembl.protein/ensembl.gene as a scalar or a list,
  and ensembl itself can be a list of dicts (zebrafish) — proteins are kept scoped to their own
  gene entry, not pooled.
• MGI double-prefix: MyGene returns MGI pre-prefixed (MGI:3704398) but the downstream consumer
  prepends MGI: itself; the handler strips a leading MGI: to avoid MGI:MGI:3704398.
• Atomic writes: each dataset is written to BioMart.tsv.part then os.replaced onto
  BioMart.tsv only after the whole taxon is pulled, so a mid-pull failure can never leave a partial
  file that a later run mistakes for complete (a reviewer reproduced that data-loss path; it is
  regression-tested).

Design

• Default unchanged on purpose: flipping ensembl_source changes gene/protein clique membership
  and should be validated with a source-impact report + clique diff on a cluster build first.
  Deferred: switching the default; a full species pull smoke-test on the cluster.
• Testable offline: every network touch is behind an injectable http_get(url, params) -> dict,
  so the parsing/paging core runs against committed real fixtures (tests/data/ensembl_mygene/,
  captured verbatim, each named by its gene id).

Docs

• Rewrites docs/sources/ENSEMBL/Download.md: documents both backends and the MyGene field
  mapping/coverage, corrects the attribute-limit description (the limit is on external-reference
  attributes), and adds the Ensembl release-116 / July-2026 legacy-platform retirement caveat.

Notes (not fixed here)

• gene.build_gene_ensembl_relationships renders prefixes as Python sets ({'NCBIGene'}:1017),
  but this is dormant: gene/concords/ENSEMBL is not in config.yaml gene_concords, so
  gene_compendia never consumes it (the live Ensembl↔NCBIGene linkage comes from NCBIGeneENSEMBL /
  gene2ensembl.gz). Pinned by a test with invert instructions; worth its own issue if it should go live.

Testing

• uv run pytest -m unit tests/datahandlers/test_ensembl_mygene.py -q → 38 passed, 2 deselected
  (includes a round-trip through the real write_ensembl_gene_ids / write_ensembl_protein_ids /
  build_gene_ensembl_relationships).
• uv run pytest -m network --network tests/datahandlers/test_ensembl_mygene.py -q → 2 passed
  (live MyGene smoke tests, incl. a real fetch_all→scroll_id handshake).
• uv run pytest -m unit -q (full CI gate) → 440 passed.
• uv run ruff check / uv run ruff format --check / uv run snakefmt --check / uv run rumdl check
  → all clean.
• uv run snakemake -c1 get_ensembl -n → DAG builds (exit 0).

Questions for the reviewer

• Default flip. Should ensembl_source: mygene ever become the default, or stay a dev/partial
  option (it cannot cover yeast/dicty)? I left it opt-in pending a cluster impact report.

---

AI-drafted: This PR was drafted by an AI coding agent (pi) at the maintainer's direction and is
pending human review. Verification commands above were run by the agent.