Fix correctness bugs in pileup alignment metrics

[micah-olivas]

Fixes four correctness bugs in the per-well pileup viewer (src/usortm/demux/streakout.py). The canvas rendering itself was fine — the problems were all in the metrics layered on top of the alignment, and because this view is what decides whether a well looks recoverable, each could push that call the wrong way.

Each was confirmed against synthetic reads before and after, and the deletion encoding was additionally verified end-to-end through a real minimap2/samtools alignment.

1. Deletions were invisible to "Read identity"

Deletions and uncovered positions were both encoded ("-", True), and identity excluded every - from the denominator:

read	before	after
30 bp deletion, zero substitutions	100.0%	96.7%
5 substitutions, full coverage	92.4%	92.4%

The cleaner-looking read was the more broken one. Now PILEUP_DEL (*) and PILEUP_NOCOV (-) are distinct, and only uncovered positions are excluded — a deletion is evidence of disagreement, not missing data.

2. The protein track silently preserved frame across a frameshift

Deleted positions were padded with N, one per reference position, so the consensus stayed reference-length and never shifted. A 1 bp deletion rendered as one changed residue:

before   refAA : KKKKKKKKKKFFFFFFFFFF
         consAA: KKXKKKKKKKFFFFFFFFFF    ← frame preserved, reads as a point substitution
after    consAA is 19 aa vs 20 aa, and residues downstream diverge

The insert is now sliced in reference coordinates first, then deleted positions are dropped, so the frame genuinely shifts. The net indel is labelled -1 bp frameshift or -3 bp in-frame indel so a shifted translation is legible rather than looking like garbage.

3. Read counts contradicted their own percentage

n_reads was the post-filter aligned-row count while frac came from the pre-filter group, so a header could read 0 reads (75%). The empty-state message interpolated that same zeroed count and therefore always printed (0 reads unaligned) — the one number it existed to convey. Sections now carry n_reads (group size) and n_aligned separately; headers read 12 of 30 reads aligned (45%) when they differ. This matters most where the midpoint filter drops concatemers, which is common and deliberate.

4. No depth floor on flagged columns

The >10% disagreement threshold was a fraction of covered reads, so one read covering one position flagged it at 100% — spurious triangles at every read end. Flagging now requires PILEUP_MIN_FLAG_DEPTH (3) covering reads. Deletions count toward disagreement.

Also fixed

• All three row builders now share one parser (_rows_from_aligned_bam), so the symbol encoding cannot drift. _build_pileup_from_bam was dead code still carrying the old encoding — exactly how this bug would return.
• Pipeline could hang. _run_pileup_alignment now closes the parent's copy of minimap2's stdout and checks both exit codes. Previously, if samtools exited early minimap2 never received EPIPE and wait() could block forever; a failed toolchain also degraded silently to an empty pileup. Verified: the samtools-missing case now returns in 0.01s.
• Consistent empty-group handling — the orient-ref branch emitted a section with no reference where _make_section skipped the group.
• Deletions render in their own colour with legend and tooltip entries; the legend now states that insertions are not shown (inherent to one column per reference position, previously undocumented).
• Canvas backing store clamped to a max area — large pileups exceeded iOS Safari's ~16.8M device-px cap and rendered blank.
• Inline JSON escapes </ so a payload cannot close the <script> block.
• Removed an unreachable JS fallback ([] is truthy, so it never ran).

Verification

• New tests/test_pileup_html.py, 30 tests — the file had zero coverage before. The render-layer tests need no aligner; the alignment-failure tests skip when minimap2/samtools are absent.
• Full suite: 202 passed.
• Real-alignment check: a 30 bp deletion through actual minimap2/samtools yields exactly 30 PILEUP_DEL positions and 96.67% identity (was 100%); perfect read 100%, 1-substitution read 99.89%.