0.21.0 - 2026-09-08

Release Notes

• Adopt the expanded-analysis implementation as the prospective performance
  baseline. Complete analysis and additional evidence can cost more than 0.20:
  the controlled Alamofire default sample measured 2.54 s versus 3.96 s.
  This accepted tradeoff preserves complete results; subsequent regressions keep
  the existing 5% and 5 ms limits. See the release qualification for scope and
  absolute latency/resource measurements.

• Compress repeated candidate relations while preserving every accepted edge,
  source constraint and connected-seed rule. Reduce query JSON copying and
  repeated navigation lookups, preserving all fields and output order.

• Share exact multiset and anchor intersections across dense comparisons, and
  select connected seeds with bounded streaming queues. Avoid redundant syntax
  budget walks and repeated normalization index invalidation.

• Keep dense accepted relations and reported-site evidence compressed through
  grouping and ranking. Rust report APIs use AcceptedEdges and explicit
  GroupEdges coordinates. Large site graphs materialize once when their edges
  are read; CLI JSON and pair-evidence policies are preserved.

• Prepare independent site mappings in parallel and batch identical exact
  evidence into sparse block masks, retaining ordered scores and witness winners.

• Reuse candidate inputs and established exact-group proofs, avoid hashing
  ineligible exact-scoring inputs, and construct evidence by group. Preserve
  alignment scores and identity bytes while reducing allocation and repeated work.

• Reuse proven exact-value candidate classes, project complete isolated rows onto
  reported sites, and encode large JSON lists in parallel without changing bytes
  or first-winner evidence rules.

• Store certified complete site relations as a site count and one exact score,
  preserving ordered edges while avoiding redundant graph construction. Carry
  uniform non-nested source relations to ranking to reuse its existing site mapping.

• Resolve enclosing contexts per file in parallel, reuse group evidence for exact
  edge classification, and reduce duplicate evidence directly for two reported sites.
  Avoid redundant file tables and per-node child copies during parameter inference;
  preserve source identities, admission rules and ordered evidence winners.

• Certify complete uniform explicit relations before expanding their edges, and
  collect small groups without nested parallel setup. Preserve source exclusions,
  exact scores and complete query output.

• Share scoring across relocated anchor evidence, retain prepared intersections
  for mostly unique inputs, and estimate sparse comparisons from both feature
  lengths. Prepare large accepted targets in parallel and union dense candidate
  neighborhoods with exact bitsets, preserving complete query output.

• Index accepted target positions only for files represented by the source row,
  and retain row structure through scoring instead of flattening and regrouping
  every relation. Reduce dense-analysis memory while preserving every target,
  ordered score, source exclusion and complete query output.

• Stop C-header admission parsing when every remaining parse necessarily contains
  a syntax error, while preserving clean-tree validation and parser reuse safety.

• Reuse exact scoring inputs and repeated score sums without changing sequential
  floating-point rounding or complete accepted-pair evidence.

• Avoid repeated Rust scope traversal for type references without a matching Tokio
  runtime import, preserving the existing type-evidence and shadowing checks.

• Preserve member navigation limits and expose source-opening commands in member
  lists; distinguish family comparison evidence from selected-member inspection.

• Inspect bounded surrounding source with member-context=N, retaining the
  original matched region and using available context space at file boundaries.

• Reopen new saved-analysis observations using captured live family handles;
  repair external-root source lookup and report actual verified/unavailable reads.

• Explain folded and displayed query populations and omit redundant recording
  guidance after successfully saving a caller review.

• Connect saved-analysis review to explicit hash-verified source inspection and separate,
  caller-owned review records with conservative applicability checks.

• Explain extraction support and scope evidence; add member facets and recovered-overlap
  explanations while preserving family identity and full-family metrics.

• Make query sorting preserve family selection and overlap representatives; stop ranking zero-invariant source matches as extractable helpers.

• Distinguish cross-language computation matches from callable reuse, propagate Rust test context into nested/cached regions, and add scope/evaluation exploration routes.

• Keep human analysis comparisons concise, with explicit observations and readable member counts; reveal capture context through full while retaining incomplete-source diagnostics.

• Capture admitted code-family analyses with query --save-analysis FILE and explore
  their changes offline with query --before FILE --after FILE: reason/evidence facets,
  before/after observations and runnable next commands. Separate versioned artifacts
  preserve coverage and ambiguity; existing query/baseline/gate contracts are unchanged.

• Add source-region snapshots and bounded, explicit cross-snapshot correspondence
  with nose regions snapshot/compare; preserve ambiguity and separate evidence
  continuity from caller-owned review decisions.

• Query JSON v10 adds nullable SHA-256 region and review-content signatures that
  survive location-only changes and preserve copy multiplicity. Existing family,
  baseline, ignore, and SARIF identities are unchanged. Non-base integrations must
  accept schema 10 (base queries remain schema 8); see docs/region-identity.md.

• Complete review keys for abstraction templates, connected/bounded regions, and
  locked external near/exact packs, including content-pinned dependency and receipt
  evidence. Correct pre-release v10 key churn from occurrence-salted analysis and
  whole-file syntax container spans. Unit/stream cache schema v5 rebuilds older
  cached features; detection and navigation identities retain their contracts.

Fixed

• Explain duplicate-member growth separately from scope and pack-fact changes in
  saved analyses; retain the conservative review-evidence checks.
• Preserve skipped-source diagnostics for offline coverage inspection, prioritize
  recheck observations, and separate total/selected/shown counts. Labeled next actions
  preserve format and view and offer explicit budget/empty-selection recovery.
• Summarize member correspondence in full analysis details and expose complete base
  candidate evidence through a context-preserving JSON action before the gate action.
• Make saved-analysis witness facet links executable and accept canonical witness
  kinds plus existing query aliases; advertise canonical values in capabilities.
• Preserve moved-region candidates when the original current function or file is
  missing, independently of unavailable semantic-change evidence.
• Show bounded original-region candidates across changed files in base= semantic
  evidence. Competing byte-identical occurrences make weak range alignment advisory;
  actual named edits retain their evidence, and the divergence gate is unchanged.
• Bound recursive oracle calls independently of execution steps, reporting a
  budget exclusion instead of overflowing the stack on nonterminating recursion.
• Accept long generated type/operator chains in pinned real-world corpora while
  retaining explicit parse-depth and candidate-work limits. Deduplicate overlapping
  LSH pairs before allocation, and avoid approximate candidate work in exact-only scans.
• Fail incomplete source discovery/reads instead of reporting a passing CI gate,
  including when reusable cache entries exist.
• Refresh watch dashboards for Markdown, ignore, and external configuration changes;
  derive replacement code snapshots and digests from the same session generation.
• Invalidate IL indexes automatically on mutation and preserve the flat serialized
  arena format. Evidence append callers can use Il::push_evidence.
• Respect explicit Markdown roots beneath vendor/build-named ancestors, and aggregate
  accepted Markdown pairs once per component instead of rescanning all pairs per family.
• Repair corrupt cache entries through replacement on every platform.
• Preserve all LSH candidate pairs in dense buckets instead of losing pairs at 49 units;
  deduplicate identical band memberships before emission. Dense buckets cost quadratic work.
• Validate detection/scoring configuration per analysis and bind cached scores to its
  effective values. Group witness construction now uses typed evidence variants while
  preserving the existing JSON fields.

Changed

• Extend the offline TypeScript oracle to primitive-key Map/Set membership and size,
  with source-checked projections and Node calibration. Value lookup, iteration, and
  mutation remain outside this projection; source-analysis cache artifacts are rebuilt.
• Add a reproducible development profiler for real-corpus clean/cold/warm equivalence,
  runtime, memory, leaf edits, and watch crash recovery.
• Centralized auxiliary development-tool versions and checksummed
  macOS/Linux assets in one checked policy, with a read-only doctor, explicit
  idempotent bootstrap, exact hosted CI consumption, drift self-tests, and
  scheduled upstream-update reporting.
• Added a checked documentation lifecycle catalog and freshness gate for every
  wiki page, with explicit current-guide, maintained-reference, decision,
  active-roadmap, and historical-record routing. Current guidance now stays on
  the primary indexes while append-only experiment, dogfooding, release, audit,
  and closeout evidence is routed through a dedicated historical index.
• Moved divergent-edit tier, taxonomy, reason, and default-gate decisions into
  a deterministic nose-detect policy boundary with focused owner tests, while
  retaining CLI JSON/SARIF, Git, exit-status, and user-workflow contracts.
• Added fail-closed, report-only pull-request change routing backed by the
  checked gate registry, historical/synthetic path matrices, auditable route
  receipts, and a stable aggregate qualification result. Main, release,
  unknown, deleted, renamed, dependency/toolchain, CI-policy, and shared
  semantic changes retain complete qualification; no gate skips are enabled.
• Split hosted workspace test compilation from execution: pull requests use a
  dev-semantic ci-test profile without unused debug data, while main,
  nightly, local-full, and release qualification retain the full optimized
  workspace suite. Optimized executable contracts now build in parallel.
• Added job-scoped, read-only hosted CI timing receipts and summaries with job,
  named-gate, fan-in wall-time limiter, runner/toolchain, and recent p50/p95
  data. Superseded commits now cancel only within the same pull request; main,
  release, scheduled, and unrelated runs retain full execution.
• Collapsed the public JavaTypeReference simple-name resolution model to its
  only supported policy: an imported, unshadowed type name. Removed
  JavaSimpleTypeResolution, made the reference coordinates private, and
  exposed read-only accessors; downstream struct literals and direct field
  access must use JavaTypeReference::imported_unshadowed and its accessors.

Install nose-cli 0.21.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/corca-ai/nose/releases/download/v0.21.0/nose-cli-installer.sh | sh

Install prebuilt binaries via Homebrew

brew install corca-ai/tap/nose

Download nose-cli 0.21.0

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