feat: per-locus CPU-time profiling (thread CPU time + harness)

[wdecoster]

Adds per-locus profiling instrumentation and an analysis harness to find which loci are consistently slow and why.

Why

The debug-only ;TIME= INFO field measured wall-clock time, which is inflated whenever other processes contend for the machine — making per-locus timings unreliable across runs.

What

• Thread CPU time. TIME= now reports thread CPU time via repeats::thread_cpu_time() (clock_gettime(CLOCK_THREAD_CPUTIME_ID), hence the libc dep). This only advances while the thread is on-CPU, so it's immune to other-process noise. Valid here because each locus is genotyped start-to-finish on a single rayon worker, so the thread-CPU delta around it is exactly that locus's cost. Stays correct under multi-threading too.
  • RepeatInterval.created is now a Duration; new cpu_elapsed() / time_field() helpers collapse the four duplicated timing blocks in vcf.rs.
• misc/profile_loci.py. Runs STRdust over N replicates, drops --warmup cold-cache runs, and aggregates per-locus median / MAD / CV CPU time, joining each locus with VCF features (depth, allele length, stdev, cluster count, outliers, QUICKREF). Prints the consistently-slowest loci and Spearman feature correlations. --from-vcfs re-aggregates existing debug VCFs without rerunning.

Findings (chr22:20–21 Mb, 1607 loci, 4 replicates)

• Measurement is reliable: slow loci show low CV (~0.06–0.15) across replicates.
• Per-locus VCF features are weak predictors of CPU time (strongest is repeat-length stdev, Spearman ~0.30; depth/length ~0.18).
• The real driver is genomic context: slow loci cluster in the 22q11.2 segmental-duplication block (~3× slower at equal depth), where reads multi-map and minimap2 spends the time on chaining/alignment candidates.

So this is a diagnostic tool, not itself a speedup — but it cleanly ruled out feature-based heuristics and the read-slicing direction.

Notes

• Branches off the 0.19.0 chore already on main.
• libc was already in the lockfile transitively; the direct dep adds a single line.