v3.16.0 — the channel vocabulary reaches the ablation machinery

3.16.0 — 2026-09-24

Added — the channel vocabulary reaches the ablation machinery

mir.explain has always asked "which named channel of the matrix carries this signal", and
mir.signature has always emitted columns whose second field names exactly such a group. Nothing
connected the two, so anyone asking the question over a signature matrix rebuilt the channel map by
hand.

• channel_spec(tier, columns=…, per_locus=…) returns a mir.explain.ChannelSpec over signature
  columns, ready for channel_report / channel_drivers. Attributability comes from the layout's
  Block declarations, so only the four geometry channels (rsig:contrast, rsig:phiv,
  rsig:phij, rsig:phic) will attribute, and the rest raise rather than answer a category error.
• CHANNELS, channel, channels, channel_table are re-exported from vdjtools.signature
  (>= 3.13.0), so a caller of mir.signature never has to know where the contract is implemented.
• mir signature --channels prints the vocabulary and reads no input, beside --describe.
• MODELS is exported — the named scale references were reachable only by string before.

examples/signature_pipeline.py gains a channels section, so the gallery's read-this-first
notebook now ends on how to read 688 columns rather than on how to produce them.

Changed — the README says signatures are an extended option, and says which scale to use

The signature section is shorter, explicitly marked extended, and links to the docs rather than
restating them. It now carries the --scale table (which corpus each reference was fitted on,
public SRA bulk RNA-seq vs deep targeted sequencing, and how many of 1,403 columns each scales)
and explains weighted vs unweighted in the terms that decide it: one vote per study against one
vote per sample. A Where to start table routes each task to its docs page.

Fixed — two stale claims in the shipped docstrings

MODELS still said blood and tissue "are not in the wheel yet"; all six have shipped since
3.15.0. And the cstar comparison quoted 0.1256 for blood TRB from a pre-3.11 fit — the reference
that ships carries 0.1072, so the amplicon-vs-RNA-seq gap on TRB is 3.8x, not 3.2x.

The README also still advertised --no-filter-functional, which has been refused since 3.12.0: a
stop codon is in seqtree's alphabet, so an unfiltered frame does not crash, it embeds to a finite
meaningless distance and contaminates the geometry silently.