Release Notes
Added
- Qualified the final v0.20.0 product tree against the checksum-verified published
v0.19.0 binary (#943). The checked release evidence
binds all-120 default/semantic/no-pack and frozen base-view query gates, 30-replay
cache and watch campaigns, and complete same-commit Soundness Lab nightly/deep
evidence; every blocking gate passes with no unresolved runtime signal, false
merge, canon violation, cache-history mismatch, or material performance regression.
- Made
--cache-dir transactional, corruption-safe, concurrent, compact, and bounded (#876).
Immutable checksummed records now commit through one complete generation pointer; interrupted
or concurrent writers cannot expose partial state, corrupt/oversized entries recompute, and
a 5 GiB default budget evicts performance data without changing results. New stable
nose cache status|prune|clear commands and --cache-max-bytes / [query].cache-max-bytes
expose storage control without deleting unrelated files.
- Added local
nose.semantic-pack-lock.v1 project authorization for typed v1
semantic packs. nose semantic-pack lock/status content-pin manifest identity,
semantic digest, selected rows/channels, dependency files, and optional receipts;
query validates configured locks before analysis, rejects stale/escaped/conflicting
decisions independent of load order, and can admit dependency-backed near rows or
receipt-backed, separately attributed external-claim exact collection factories.
v0 and unlocked v1 remain metadata-only. External packs are local, explicit opt-ins,
disabled by default, and data-only: nose adds no provider-code execution, registry,
network resolver, installer, or parser plugin. A shipped Vavr 0.9.x List.of
reference pack demonstrates the complete lock, receipt, attribution, and rollback
workflow without claiming builtin certification.
- Closed the #847 divergent-edit precision-first cycle without overstating readiness.
Direct changed-to-skipped targets, bounded semantic-change evidence, and pair-local
variant signals now give users more precise review context, while one internal policy
decision keeps human, JSON, SARIF, and exit status aligned. No admissible v3 policy had
development support, so the blind population remains sealed and v2 stays opt-in. The
official-v0.19.0 closeout is output-compatible after removing reviewed additive evidence
but misses the runtime gate at +8.74% control-adjusted; default-on enforcement is not
recommended.
- Added fail-closed Soundness Lab PR, nightly, deep-campaign, and 0.20 release
gates. Nightly evidence covers every pinned repository exactly once, retains artifacts on
failure, binds shards to one source commit and binary, fully diffs non-blocking advisory
disagreements against the published v0.19.0 binary, and rejects missing artifacts, changed
pins, timeouts, false merges, canon violations, coverage regressions, unregistered claims,
unguarded Tier-A cells, and unattributed exclusions. Deterministic shard evidence is separated
from diagnostic logs, release failures retain a summary, and the release commit's product tree
must match the checked binding.
- Kept mixed-exit fragments distinct from whole functions in exact fingerprints. A conditional
fragment that may return or throw but can also fall through no longer collides with a whole
function whose completion has different enclosing-control behavior; the Soundness Lab found
and reproduced this boundary across pinned repositories. The audit oracle now distinguishes
explicit throw from ordinary evaluation errors independently of the product classifier, and
cache schema v12 prevents pre-fix feature entries from being reused.
- Committed the #846 held-out review inputs without publishing their source. The
official v0.19.0 replay reproduces 54 queries, 1,564 candidate commitments, and the
exact 214-family sealed selection, then creates three persona-specific packets outside
Git. Only secret-nonce whole-packet hashes are public before voting. Three independent
reviews rejected two earlier drafts—the second was still source-fingerprintable
214/214—so the accepted contract explicitly promises procedural product-metadata
blindness, not identity hiding from a reviewer who searches public source. All three
votes must freeze together and blind-ID arbitration must freeze before reveal.
- Closed the #845 dev-only residual-ranking experiment with a fully judged no-go. Three
independent blind reviewers labeled all 219 frozen frontier families; exact tuple
agreement covered 129 and an independent arbiter resolved all 90 worthiness-or-reason
disagreements. The exact-key-only overlay gives every one of 46 formulas 100% top-10
truth coverage on every repository. Current is 387/658 (58.81%); the best
coverage-guarded formula is 449/658 (68.24%), 12 hits short of 70%, and leaves C at
44/90 (48.89%), one hit below the language floor. Repository-CV out-of-fold is
416/658 (63.22%). No proposal or signal is retained, held-out stays closed, and product
code, ranking, surfaces, family membership, and worthy recall remain unchanged.
- Froze the judgment-blind #845 residual-ranking calibration and complete dev top-up
frontier. The full 29,348-family universe evaluates 46 unique transparent formulas,
rejects ambiguous overlap labels, uses exact reported-position gates and deterministic
total orders, and remains explicitly evidence-incomplete rather than treating missing
labels as product failures. A non-adaptive 219-family panel packet covers every
unresolved top-10 position of every formula and therefore every repository CV fold.
The packet contains no judgments; held-out remains closed and product behavior is
unchanged until the separately committed panel result decides go or no-go.
- Recorded the checked #844 proof/actionability no-go. Three independent source reviews
agree that removing the proven-channel protection reaches only 60% non-action precision
on its direct frozen boundary (0% for trivial, 75% for shallow-extraction), while
the broader exact/subdag cohort reaches 42.19% and contains 37 worthy helper,
parameterization, data-table, and base-extraction hard negatives. Each reviewer's
five-judgment record is source-packet bound, and future admission requires a 90% point
precision and one-sided 95% Wilson lower-bound gate. CI reconstructs the no-go and zero
product drift from the #841 taxonomy, normalized executable-code identity, and the
checked #843 behavior, quality, and official-v0.19.0 performance evidence.
- Classified strict declaration-only type contracts as reason-coded
declaration output
(#843). The all-member rule consumes existing language-neutral origin facets for Java,
TypeScript, Rust, and Swift; incomplete, narrowed, runtime/data/implementation,
default-body, extension, enum, schema, and unknown evidence stays fail-open. All five
frozen positives move, eight worthy hard negatives stay default, 54,754 dev families
and their order remain stable, and full-universe worthy recall remains 2716/2849 with
zero regressions. Among currently labeled default positions, interim dev P@10 rises
from 58.05% to 59.04% while the worthy-hit count stays 382; replacement-label coverage
is 98.33% for #845/#846 to close before a final precision claim. Existing frontend
facets now truthfully mark Java field initializers, Rust defaults/macros/attributes,
and parser-recovered Swift bodies. The official-v0.19.0 all-dev result is +0.42% raw /
+0.35% adjusted; the exact 66 -> 25 -> 6 -> 2 repository escalation across
3 -> 9 -> 21 -> 40 iterations ends -0.69% adjusted with no material regression.
- Classified source-coherent Jazzy documentation as reason-coded generated output (#842).
Every HTML member must carry both a Jazzy asset and an Apple/Dash symbol anchor within
a bounded 64 KiB head; missing, partial, unreadable, and hand-written lookalikes stay on
their ranked surface. All 30 frozen head/deep positives move to generated, the three
worthy HTML hard negatives remain default, and all top=0 preserves ordered family
IDs and every non-surface field. Default-only slices also remain visible when a generated
primary leaves that view, including equivalent and negated surface filters. Only
Alamofire changes across 66 dev repositories. The official-v0.19.0 all-dev comparison
is -1.26% raw / -1.72% control-adjusted; the exact 3 -> 9 -> 21 -> 40 focused chain
finishes -1.02% adjusted, with no material aggregate, repository, or stage regression.
CI reconstructs cohort/corpus summaries from bound per-repository hashes and recomputes
every performance escalation edge from raw reports.
- Added the complete #841 dev default-head failure taxonomy and bounded-lever contract.
The published v0.19.0 output is re-bound at all 658 positions to full raw families,
source bounds, truth and disjoint mechanical buckets, origin/generated/ranking/path
cross-tabs, and rejected heuristics. Three independent source audits each confirmed
all 20 Jazzy-generated and four strict declaration-only deep positives as
non-actionable, while worthy HTML, incomplete/missing origin, reusable-body, and 35
proof-backed refactorings remain hash-bound hard negatives. Generated provenance and
strict declaration-only type contracts clear the 90% gate for #842/#843; blanket
exact/subdag exemption removal is frozen as a #844 no-go. Held-out source remains
closed, and CI reconstructs the compact decision overlay from its checked dev-only
label projection, 8.1 MB core, standalone truth-free audit packets, and all three raw
vote files. Exact nested schemas and #840/live source bindings replay raw-derived
facets, predicates, buckets, frozen deep/audit cohorts, bounded generator evidence,
and audit-key uniqueness rather than trusting dependent hashes. Frozen v5-dev and
generated-evidence/packet-set commitments, exact command provenance, and a
source-locked one-time vote rebind reject consistently rehashed substitutions and
synthetic prior judgments while keeping ordinary CI independent of ignored corpus
checkouts.
- Added the split-safe v7 default-head label runway (#840): every v6-unmatched
dev top-10 family plus deterministic rank 11–30 repository samples now have
independent three-persona votes and explicit arbitration, while a pre-hashed
held-out selection exposes neither source nor judgment. The precision-only
overlay raises dev label coverage from 66.41% to 100%, preserves byte-frozen
v5/v6 and worthy recall, and reveals the complete 58.05% dev default-head
baseline for the bounded #841 failure taxonomy. Exact held-out schemas reject
unknown or value-encoded source/judgment data and held-out overlays. Repository
identity, commits, and match counts are bound to the hashed corpus and candidate
commitments. The exact manifest binds its byte and family projection to frozen
v6, CI replays the complete raw-vote-to-component evidence chain and verifies
the evaluation sidecar, and split-local bootstrap streams keep unchanged
held-out intervals independent from dev label growth.
- Split the product-quality evaluator's user-facing default-head precision from
full-universe worthy recall (#839). The published v0.19.0 binary now anchors a
hash-identified 120-repository baseline; executable default-list ID-and-order
and complete bare-dashboard top-five/summary parity, uncached measurement,
per-surface counts, explicit
historical all compatibility mode, and deterministic reports prevent future
ranking work from measuring the wrong query surface.
Install nose-cli 0.20.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/corca-ai/nose/releases/download/v0.20.0/nose-cli-installer.sh | sh
Install prebuilt binaries via Homebrew
brew install corca-ai/tap/nose
Download nose-cli 0.20.0
