Disable --pairguide auto and validate extraction windows up front

[davidliwei]

Disables --pairguide auto, which has never produced a correct result, and closes the silent-failure paths that removing it exposes.

Why auto goes

It reused mageckcount_search_variable_region, which walks read columns and stops at the first one whose most common base exceeds 60%. That finds a UMI — a uniformly random run. A second guide is drawn from a fixed library, whose per-column base composition is lumpy by construction, so the search is structurally wrong for guides: no threshold makes a library column look uniform.

Measured before this change, with explicit windows as the control:

Layout	auto	explicit --pg-start*
R1 = guide1+guide2, 8-guide library	window 0-4, 0 rows	8 rows
R1 = guide1+constant, R2 = guide2	window 0-4, 0 rows	8 rows
same, 200-guide library	window 0--1, 0 rows	200 rows
R2 = guide2, nothing after guide1 on R1	uncaught ValueError	4 rows

Each failure exited 0 with a header-only pg_count.txt — the confusion reported in #15. Any pipeline pinned to auto today is already producing empty output, so rejecting the choice turns a silent wrong answer into a loud one. #32 tracks reimplementing it on mageckcount_trim5_auto, which already locates guide 1 by counting library matches and exits non-zero when almost nothing matches.

The path it leaves behind was also unguarded

An explicit window is now the only option, so omitting one has to fail cleanly. It did not:

--pairguide secondpair, no --pg-start-2/--pg-end-2   before: rc=0, 0 rows, one ERROR per read
                                                      after: rc=255, one ERROR naming both options

The check lived inside the per-read loop — one log line per read on a real FASTQ — and then sliced an empty string and finished successfully. It now runs once in mageckcount_check_extraction_window, before any FASTQ is opened, and covers --umi too, which shares the code path and the bug.

Also fixed in the shared search, still used by --umi auto

• A failed search was accepted as success. The second-read result was taken when either bound was set (>= 0 or >= 0; the first-read test at the line above correctly uses and). The failure value (0, -1) then sliced each read as [0:-1] — the whole read but its last base — so every read yielded a distinct "UMI". A run with an 8bp UMI on read 1 filled umi_count.txt with 2,000 rows built from unrelated read-2 sequence. This is the worst of the group: nothing about the output looks wrong.
• max() over an empty column table raised an uncaught ValueError when nothing follows the guide on read 1 — the normal layout for --pairguide secondpair.
• An unreachable --pairguide branch in mageckcount_processonefile: args.umi is assigned from args.pairguide before it runs, so the preceding test always won. Had it run, it set args.umi='none' and disabled pair extraction entirely — its being dead is the only reason auto got as far as it did.

Tests

Six new, each written failing first. 38 pass total.

test_pairguide_auto_is_rejected · test_pairguide_secondpair_requires_its_window · test_pairguide_firstpair_requires_its_window · test_umi_secondpair_requires_its_window · test_umi_auto_fails_loudly_when_no_variable_region_is_found · test_umi_auto_reports_reads_with_nothing_after_the_guide

The window tests assert the error appears once, not once per read.

Not covered here

--umi auto still fails to find a UMI that runs to the end of the sequence: var_end is never assigned in that case and the search returns (start, -1). This PR makes that fail loudly instead of silently; actually finding the region is filed separately. The remaining --pairguide test coverage and the ctab_umi[None] accumulation stay in #29.

Refs #29, #32.

[davidliwei]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. You're on a roll.

Reviewed commit: 7e21c73eb2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".