openvax / topiary / 34801412262-1

92%

push

github

Dedicated CachedPredictorCoverageError; diagnostic and efficiency fixes (#308)

An ultra code review of 5.55.0/5.55.1's changes (#302, #300, #304)
found several real issues alongside some already-considered and
declined ones. This addresses the confirmed, low-risk findings and
files follow-up issues for the two that are real design questions
rather than bugs.

Fixed, each verified independently before the fix:

- CachedPredictor raised bare KeyError for its two intentional
  coverage failures. main()'s handler caught it, which meant an
  unrelated dict-lookup bug anywhere in predict_epitopes_from_args'
  call graph would silently become a clean CLI error instead of a
  traceback. Added CachedPredictorCoverageError(KeyError) as the
  specific, exported type both raise sites use; the CLI now catches
  that instead of bare KeyError. Existing `except KeyError` callers
  are unaffected since it subclasses KeyError.

- _raise_on_uncovered_occurrences reported only the first mismatched
  flank context when a cache held more than one for the same
  (peptide, allele, kind, allele_set) - reproduced directly with two
  genuinely distinct cached contexts, only one of which the old
  message named. Now lists every stored context (capped at 5, with an
  accurate "(and N more)" past the cap driven by a true count kept
  independent of the cap) instead of silently dropping the rest.
  unmatched_context's accumulation is capped for the same reason: nothing
  past the first few rows was ever read.

- _flanks_apply_at sliced and upper-cased the whole remaining sequence
  on every (cached row, occurrence) pair, when only the cached flank's
  own length can ever match. Bounded the slice to that length;
  verified equivalent to the old logic across 20,000 randomized trials
  before relying on it, plus the existing edge-case tests (shorter
  available context, flank longer than available context).

- A bracket after an identifier the parser doesn't recognize as a
  kind now... (continued)

42 of 44 new or added lines in 5 files covered. (95.45%)

7123 of 7731 relevant lines covered (92.14%)

0.92 hits per line