fix(error_models): catch mis-stated binned-qual anchors (#263)

[cjfields]

Closes #263.

The problem, measured on real data

trans is indexed by each unique's mean quality, so binned reads do not produce a cleanly binned matrix. They leave mass spikes at the true bins and a thin smear of averaged qualities between them. On 5 NovaSeq 16S samples: Q37 70%, Q36 11.7%, Q35 5.5%, Q25 3.4%, Q11 1.7%, and every other column under 1%.

An anchor that misses the true bins lands in the smear. It still fits, but from a handful of transitions: anchor Q12 rested on 37 transitions, against 1.66% of all transitions at Q11 next to it. With 2,12,23,37 on that data, 241 of 456 off-diagonal rates sat at the 1e-7 floor and 21 at the 0.25 clamp, up to 6 orders of magnitude from the correct model. R builds the identical model (1.9e-15), with no error.

Changes (src/error_models.rs, check_binned_anchors)

Case	Before	After
data outside the anchors	error (as R)	unchanged
observed min/max not an anchor	warning (as R), repeated every iteration	same warning, printed once per run
anchor inside the observed range with no observations	rates across the gap silently 1e-7; the whole model when every segment is affected	error, naming the empty anchors and the qualities present (R returns NA, or stops)
anchor holding less mass than a non-anchor quality within ±2	silent	stern warning, once per run, with both masses: "THE ERROR MODEL IS LIKELY WRONG"
anchor below or above the observed range (NovaSeq's Q2 after trimming)	silent, flat-filled	unchanged, and now pinned by a test

Alternative bin schemes still fit; the warning only says why they are suspect. --errfun external remains the path for deliberately custom models. docs/commands/learn-errors.md explains the mean-quality smear and the new error and warning.

Verification

Real data, NovaSeq 16S (5 samples, filtered once, and the same files given to R and to dada2-rs):

Anchors	dada2-rs warnings	Model
2,11,25,37 (correct, unobserved Q2)	none	byte-identical to before the change, and to R within 2e-15 (F and R reads, identical sum(trans))
2,12,23,37	3, once each: min-not-anchor, Q12 vs Q11, Q23 vs Q25	builds, same as R
11,24,37 (extremes on anchors, so R's warning stays silent)	1: Q24 (0.36%) vs Q25 (3.40%)	builds

No false alarm on MiSeq i100 (bins 12,24,38, 3 samples).

Unit tests:

• an empty interior anchor errors, both in the full collapse and in a one-sided gap;
• an unobserved outer anchor is silent and gives the same model as without it;
• correct anchors on smeared data are silent;
• shifted anchors are named.

Mutation-checked: with the empty-anchor error disabled, both error tests fail.

Full suite: cargo test --profile test-opt passes on default and --features wfa (465 passed); clippy is clean on both; mkdocs build --strict passes.

A correction to the issue

#263 was filed from a synthetic case where the bin columns hold all the mass, and it says the model "collapses to min_error_rate". On real data, wrong anchors corrupt the model rather than zeroing it: about half the rates at the floor and some at the clamp. The issue text and PR #265's comments are corrected to say so.

🤖 Generated with Claude Code