chore(rust): replace LEECH_DISABLE_RUST with --backend on data prepare

[jayhesselberth]

Closes #177.

LEECH_DISABLE_RUST was a temporary workaround for #176, added because "Rust is available" and "Rust is faster here" are different claims and nothing could measure the second. They came apart badly — 13 reads/s against 130–234 on the Python pool. #176 is fixed (#178: concurrent batch dispatch, the process-global POD5 reader cache, and py.detach over the I/O), so the hatch has outlived its reason.

Why this isn't a straight delete

#177 offered three options. (a) "delete it" turned out to need a piece of (b) first: --backend existed only on predict, so the env var was the only way to force the Python path on prepare — and docs/troubleshooting.md told people to use it for exactly that, in both the throughput and read-yield sections. Removing it would have taken the diagnostic with it.

So the hatch moves to the seam #169 built rather than disappearing.

Changes

• data prepare gains --backend auto|rust|python, matching predict.
• One resolver, preparation.parallel._select_prepare_backend(), ANDs the choice with availability and rust_prepare_unsupported_reason(config). One place decides, so the startup log line and the dispatch cannot disagree.
• Scoped to prepare. The old variable raised inside the try: in _rust_accel.py, so it killed leech_core process-wide — move-table parsing and inference included, where the extension is a real win. That was far broader than the one step it was meant to switch.
• --backend rust raises instead of falling back, whether the cause is a missing leech_core, an unsupported config, or --workers 1 (the sequential path is Python-only). A forced run that quietly took the other path measures nothing, which is the whole reason the switch exists.
• LEECH_DISABLE_RUST and the DISABLE_RUST branch are gone from _rust_accel.py (the os import with them); both docs references now say --backend python.

The two backends produce identical chunks — that is what tests/test_backend_parity.py enforces field by field — so this flag only ever changes throughput. The help text says so.

Tests

Seven new cases in tests/test_prepare_dispatch.py::TestBackendSelection, covering each choice against available/unavailable and supported/unsupported configs, plus the raise-don't-fall-back property. They need no POD5/BAM fixtures, matching the rest of that file.

1015 passed, 27 skipped. ruff format --check, ruff check, and ty check clean.

Downstream

rnabioco/escapepod-models — workflow/rules/charging.smk, rule charging_leech_prepare — sets LEECH_DISABLE_RUST=1. That should now be dropped outright, since #176 is fixed and the default path is the fast one. If a pin is still wanted, it is --backend python.

🤖 Generated with Claude Code