0.13.0 - 2026-06-18

Release Notes

Adds a Swift language frontend and brings Markdown near-duplicate prose detection into
nose query as a new domain (alongside code, CSS, and HTML), plus a Type-4 soundness fix and a
higher cross-language Type-4 coverage floor. Detection of the existing languages and the
--format json query-JSON v2 contract are unchanged.

Added

• Swift language frontend (#442). Swift joins the first-party languages (Python, JavaScript,
  TypeScript, Go, Rust, Java, C, Ruby) with tree-sitter lowering and semantic coverage, so Swift
  copy-paste, renamed, and Type-4 same-logic clones are detected and proven like the others.
• Markdown same-language near-duplicate prose as a nose query domain (epic #435). nose query now reports near-duplicate prose across Markdown documents alongside code clones. Per
  capabilities over features, duplication has one entry point, so markdown is surfaced
  through nose query exactly as the CSS/HTML declarative track is — a "markdown near-duplicates"
  section in the human dashboard and a top-level markdown array under --format json — not as a
  separate command. Prose is not code, so it uses a deliberately separate nose-markdown engine:
  a character-n-gram pipeline (MinHash-LSH + winnowing + containment candidate generation →
  IDF-weighted TF-IDF verify/rank → line-level Smith-Waterman span witness), not the value-graph
  IL. Reports ranked families with a relation tier + score, an exact span witness, and
  orthogonal evidence (commonness/DF, removable lines, files). Honesty contract:
  "near-duplicate (score + witness + commonness)", never "same meaning" or "worth removing" —
  boilerplate copies are true duplicates surfaced with high commonness, never suppressed.
  Same-language only (cross-lingual/paraphrase need an LLM, out of scope). Deterministic
  (byte-identical output). Measured against frozen, LLM-built goldens (bench/markdown/, no
  human in the loop: 3 heterogeneous judges, Fleiss κ 0.70/0.71, anchor self-calibration 1.0):
  PR-AUC 0.995 / R@P95 0.96 on the code-of-conduct corpus, and an honest multi-genre baseline of
  PR-AUC 0.944 / R@P95 0.74 across 5 doc genres, candidate-recall 1.0. Built on the
  algorithm survey in docs/markdown-dup-detection-algorithm-survey-2026-06-18.md. See
  docs/markdown-duplication.md.
  (#435 #436 #437 #438 #439 #440 #441 #444)
  • Field-evaluation precision/usefulness fixes (#443, #447). (P0) default vendor-dir excludes
    (node_modules, vendor, target, …) + nose.toml exclude globs, so a non-git project's
    node_modules no longer floods the report (one field project went 145 noise families → 0); a
    min-shared-grams match-substance floor drops thin overlaps. (P1) strip GFM table
    scaffolding (pipes/separator rows are format, not content); strong-edge clustering (weak
    edges corroborate but don't chain mega-families); confidence-weighted ranking. (P2) large
    multi-file clusters are reported as templated sections ("skeleton repeated N× across M
    files"), not a clone family — so a templated-doc blob no longer masquerades as one family.
  • Containment-witness gate (P3, #449). A size-disparate small-in-large containment match
    must now have a real contiguous shared-line block, not just scattered char-gram overlap — a
    small generic stub no longer gets pulled into a large unrelated doc by common-morpheme
    coincidence. Surgical: one field project's spurious mixed-granularity families dropped 8 → 5
    with genuine small-in-large and reworded same-size near-dups preserved.
  • Synthetic recall-vs-edit-ratio benchmark + recall gate (#443, #450). The golden gates
    precision; this adds the missing recall gate — a deterministic, self-contained benchmark
    (nose-markdown::synth, committed base corpus) that injects controlled edits at known ratios
    and asserts recall floors (1.00 / 0.95 / 0.85 / 0.65 at 0/0.1/0.2/0.35 edit ratio), so a
    future change that silently sacrifices recall fails CI.
  • Multi-domain precision golden + regression gate (#443, #454). The original precision
    golden was single-genre boilerplate (Contributor Covenant CoC), which over-stated precision.
    Added a frozen multi-domain golden — bench/markdown/corpus-docs/ (165 files across 5
    genres: CLI reference, function/API reference, guides, framework docs, READMEs) +
    golden.docs.v1.json, built the same no-human way (Fleiss κ 0.71), wired into a precision
    regression gate (eval::docs_golden_precision_floor). The golden-build scripts now take paths
    so any corpus can be golden'd.

Changed

• Swift Type-4 exact coverage parity slice. Swift now has evidence-backed exact-query
  coverage for five previously-open matrix cells: typed flatMap vs nested append builders,
  module import identity, Dictionary[key, default:] map-default lookup, literal Int
  clamp ternaries vs min(max(...)), and integer-gated total-order comparison absorption.
  The implementation is deliberately proof-bounded: wrong module/member/key/default
  coordinates, optional ?? defaulting, Double clamp forms, and overloaded/String
  comparisons remain adjacent hard negatives. The checked-in Type-4 matrix moves Swift from
  15/24 to 20/24 applicable cells, and focused probes now run through
  nose query ... witness=exact rather than the deprecated scan path.
• Type-4 coverage floor raised to ≥50%. The checked-in coverage matrix
  (bench/type4/coverage_matrix.v1.json) now keeps every primary language at or above 56% of
  covered applicable cells, backed by new evidence-carrying probes (extract-method-inline,
  numeric tail recursion, …) for C, Go, and others.

Fixed

• structural_fold reassociation gated to integer heads (#434). A Lean obligation audit (all
  24) found the right-fold→left-fold loop rewrite — proven sound over Int only — could admit a
  float-valued head via the coarse ValueDomain::Number, and float +/* is non-associative, so
  results could change. Added head_possibly_float (rejects float literal / Op::TrueDiv /
  float-typed param), measured byte-identical on all 135 corpus repos (0 recall cost,
  soundness gain), plus added Lean theorems for previously-unproven firing rewrites
  (eq_commutes/ne_commutes, guard merges, and a structural_fold float counterexample).
• Swift implicit-member lowering gap (#452, #455). Swift implicit-member expressions now lower
  correctly, closing a coverage gap.

Install nose-cli 0.13.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/corca-ai/nose/releases/download/v0.13.0/nose-cli-installer.sh | sh

Install prebuilt binaries via Homebrew

brew install corca-ai/tap/nose

Download nose-cli 0.13.0

File	Platform	Checksum
nose-cli-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
nose-cli-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
nose-cli-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
nose-cli-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum