Dependency discussion: interval-overlap crates (superintervals / coitrees / rust-lapper)

Survey item from August 2026, filed so the answer is on record. Expected outcome is decline
unless a profile says otherwise.

Candidates

The September 2025 interval benchmark puts superintervals fastest or tied across cases (1.25–1.44x
over COITrees as the polars-bio default), and notes rust-lapper collapsing ~25x on adversarial
datasets, which makes it risky as a general-purpose choice.

Why the default answer is no

Annotation and junction lookups here mirror STAR's own structures rather than being free choices:
src/quant/transcriptome.rs:151 builds an explicit sorted order for binary_search, and the
junction side is keyed maps plus STAR-shaped filtering (src/junction/). Swapping in a general
interval library would not just change speed, it would change what "overlap" means at the edges
(half-open vs closed, tie handling), which is exactly where faithfulness lives.

Two further blockers before any use: both superintervals and coitrees show a non-standard
license on crates.io, which must be resolved explicitly, and neither has been shown to be on a
hot path in this codebase.

Checklist

• Profile whether interval/annotation lookup is measurably hot at all (blocked on the benchmark
  harness)
• If it is: identify the exact query shape before shopping for a library
• Resolve the license question before any prototype, not after
• Otherwise record the decline with these reasons

Start with src/quant/transcriptome.rs:151 and the junction code under src/junction/ to understand the existing lookup semantics. Profile interval and annotation lookup using the benchmark harness when available, then identify the exact query shape and check candidate licenses before considering a prototype. Done means the benchmark evidence and license status support a decision, or the decline is recorded with the stated reasons.