v0.7.4 — Sei's numbers change

Chorus v0.7.4

Sei's numbers change in this release. Its variant effects were missing a correction the upstream
authors specify, and its background covered 40 of the 21,947 tracks it predicts. If you have published
Sei results from 0.7.3 or earlier, they were computed without the nucleosome-occupancy adjustment and
against a 40-row null. Nothing else moves: no other oracle's predictions, percentiles or committed
examples are affected.

Fixed

• Sei's nucleosome-occupancy correction was absent in two independent ways
  (#224,
  #225). SeiNormalizer.normalize computed
  sum_alt from the reference array, so both scaling factors were exactly 1.0 — a no-op. And it was
  never called: sei.py built the normalizer, required it non-None, and never invoked it. The
  correction is described in its own docstring as "recommended by Sei authors" and is
  sc_hnorm_varianteffect in FunctionLab/sei-framework; normalize is now a faithful port, asserted
  with np.array_equal against a transcription rather than merely close. Measured on three real
  variants, applying it moves every one of the 40 sequence classes, by 12–48% of the variant's
  largest effect, and changes the top-ranked class for 2 of 3. At chr11:5247500 the correction is larger
  than the signal (median |Δ| 0.0181 vs median |effect| 0.0170).

  The correction is defined over a ref/alt pair, which a per-sequence _predict cannot express, so
  OracleBase gained _predict_alleles — a seam whose default is exactly the per-allele loop it
  replaced, leaving the other eight oracles untouched. SeiNormalizer.equalize generalises the pair to
  n alleles and is bit-for-bit upstream for ref + 1 alt, so chorus's result schema stays identical
  across oracles.

• 21,907 of Sei's 21,947 tracks produced no raw_score
  (#231). They classify as histone_marks /
  tf_binding / chromatin_accessibility, whose configs carry window_bp of 2001/501/501 because
  those layers were built for binned oracles — and windowing Sei's 1-element output returns None.
  score_track_effect already had a "Full output scoring (LentiMPRA, Sei, etc.)" branch that Sei's
  profiles never reached. Scoring now routes on the data shape, so a track carrying one value takes that
  branch whatever its layer. Provably scoped: the only tracks affected are scalar tracks in a windowed
  layer, which is Sei's profiles alone.

• chorus/analysis/discovery.py raised AttributeError for Sei and skipped two oracles entirely
  (#227,
  #229). It called oracle.get_all_assay_ids(),
  which Sei does not have (only a private _get_all_assay_ids), for an oracle listed in its own
  multi-track set. alphagenome_pt and cherimoya were in neither oracle set, so both returned
  {"error": "Unsupported oracle"} despite being fully implemented. Cherimoya now iterates its 407
  biosamples rather than all 1,518 experiments — measured 15.3 s per model, so 1.7 h instead of 6.5 h —
  with both counts logged rather than silently capped.

• epinformerseq.list_cell_types() returned 1 of 11 cell types
  (#227) — the instance's own cell rather than the
  oracle's catalogue, while its background carries all 33 rows. LegNet fixed this exact bug and the fix
  was never ported.

• MCP list_tracks answered for Sei from a stale literal and for ChromBPNet with models that do not
  exist (#230,
  #231). The Sei branch returned a hardcoded
  ["sequence-class"] without ever consulting the oracle — false for an oracle with 1,176 assay types.
  ChromBPNet reported 172 cell types × 240 TFs, implying 41,280 available models against 744
  that exist. Both derive from describe_tracks() now, keeping their payload shapes; ChromBPNet states
  num_chip_models and rules out the cross-product reading in words. LegNet and EPInformer-seq no
  longer hold third copies of their constants.

• A third-party timeout could fail the browser job
  (#228). igv.js reports a resource timeout in its
  own words ("IGV error: Timed out") as well as Chromium's ERR_TIMED_OUT, and one unrecognised console
  error disqualifies the whole outage skip — so a UCSC timeout reddened CI on a change that touched no
  browser code.

Changed

• Sei's background covers all 21,947 tracks, not 40
  (#226). The 40-only scope was never a recorded
  decision: the builder's docstring asserted "Sei outputs 40 regulatory classes" and everything
  downstream inherited it. The network emits 21,907 chromatin profiles that a projection aggregates into
  the 40, predict() accepts profile ids, and it returned real values whose percentile was always
  None. Compute was already being spent — project() discarded the profiles — so the cost is storage:
  2.8 MB → 1.5 GB. The builder imports SeiNormalizer rather than carrying a second copy of the
  arithmetic, and writes ids in the oracle's own TA#/CA# forms (verified to match
  _target_assays_ids() | _class_assays_ids() exactly, 0 divergence, before the build ran).

  All 1,176 of Sei's assay-type strings previously classified as other, which would have shipped
  21,907 built, verified, unscoreable rows — §8's documented failure mode at 21,907× scale.
  classify_track_layer routes profiles through the existing classify_chip_layer, giving 10,310
  tf_binding, 9,225 histone_marks, 2,372 chromatin_accessibility and 40
  regulatory_classification, zero unclassified.

  The effect null is histone-equalised and the activity null is not, because the correction is defined
  over an allele pair and a baseline is one sequence. Recorded in
  docs/BACKGROUND_NULL_PROTOCOL.md's decision log.

• The pinned backgrounds revision is now backgrounds-2026-08-16-sei-full
  (#231), matching the rebuilt artefact.

Added

• OracleBase.describe_tracks() — one call every oracle answers
  (#227). Previously this took four different calls:
  get_all_assay_ids() on four oracles, list_tracks() on cherimoya, a private _get_all_assay_ids()
  on Sei, and nothing at all on chrombpnet, legnet or epinformerseq — so a contributor adding a tenth
  oracle had no single method to implement, and consumers grew per-oracle branches that drifted from the
  oracles they described. Returns TrackRecord objects (chorus/core/tracks.py) whose track_id is
  exactly what predict() accepts, plus assay, cell type, description, a has_background flag and
  per-oracle extras.

  Concrete rather than abstract, deliberately: seven test-double subclasses across five test files
  implement exactly the previous abstract set, and a ninth abstract method breaks every one at
  instantiation. All nine oracles implement it, each count matching what is published — enformer 5,313,
  borzoi 7,611, alphagenome and alphagenome_pt 5,168, cherimoya 1,518, sei 21,947, chrombpnet 753,
  legnet 3, epinformerseq 33 — and each works before load_pretrained_model().

• Near-ties at the top-N cutoff are now reported
  (#235). discover_variant_effects returns
  near_ties_at_cutoff, populated whenever the rank N / N+1 boundary in a layer is closer than 1% —
  naming the kept and dropped track, the relative gap, and that which one is reported is not stable
  between runs. Also logged as a warning. Selection is unchanged; this only stops a coin-flip being
  presented as a result. It exists because the measured rank-12↔13 gap in tf_binding was 4.25%
  against a 4.29% worst-case cross-process drift, which is the one way the accepted drift can still
  change a conclusion while no number moves visibly.

Documentation

• Audit coverage reached 18 of 18 checklist sections
  (#223): GPU detection, per-track CDFs,
  reproducibility and scientific determinism were the four never previously exercised. Every §4 figure
  matched its published value; §3's documented trap (a bare probe reports 0 GPUs for both TensorFlow
  envs, because their CUDA libs are only on the runner's LD_LIBRARY_PATH) caught the auditor as
  designed.

Known limitations

• Cross-process drift is accepted rather than eliminated. TF_DETERMINISTIC_OPS=1 was measured to make Enformer bit-exact (4/4 runs, against a control failing 2/2 at 3.4% and 1.4%), but adopting it would require rebuilding the Enformer and ChromBPNet nulls and changing published numbers for both. Median drift is 0.016% and the published quantile_score moves at most 0.69%, so the trade is not worth it. The residual risk is which track gets reported, not the numbers: discovery.py cuts hard at top_n_per_layer and the rank-12↔13 gap can be smaller than the drift. Near-tie reporting is the follow-up. See docs/BACKGROUND_NULL_PROTOCOL.md's decision log, and audits/2026-08-14_f8_localisation.md for the measurements (#233) and the per-oracle coverage map (#234) — cherimoya, legnet and epinformerseq are bit-exact cross-process; chrombpnet and sei remain unmeasured because the gate cannot target them.

• The committed-example staleness check reports rather than fails, following that decision. It compares each
  committed example against the last semantic change to scorers.py, so this release's scoring fix
  fires it. A regen sweep is currently the wrong response: committed examples cover alphagenome,
  chrombpnet and enformer and no Sei, and the fix can only touch scalar tracks in a windowed layer,
  so no committed number can move — while F8 (audits/2026-08-14_f8_localisation.md) remains open, so
  regenerating would change Enformer's numbers for unrelated reasons. Integration-marked so it never
  blocks CI.

• CI runs the fast and browser suites only. Three real defects in this release existed solely in the
  integration suite, which needs a GPU and the per-oracle envs. Integration has to be run locally before
  a release; audits/AUDIT_CHECKLIST.md is the runbook.