v0.5.0

• Normalize Git snapshots and Python sdist/wheel layouts before diffing.
• Extract changed source from archives and show its line-level delta.
• Preserve composite findings and rank evidence across the complete change.
• Recognize prerelease transitions and apply a tight behavioral budget.
• Detect public APIs backed by floating young dependencies.
• Detect auto-loaded source or scripts that download, write, and execute payloads.
• Improve npm entrypoint discovery and local LLM guidance.
• Detect 132/132 corpus attacks with no false positives across 98 clean transitions.

Full Changelog: v0.4.1...v0.5.0

v0.4.1

Version comparison

• Four-part versions are now parsed and compared in full.
• Bump::classify rewritten as a component-wise walk to the deeper of the two versions: the first differing component decides.

Build & CI

• cargo test --locked --lib → cargo test --locked across native, FreeBSD, and OpenBSD jobs; isomer is a binary crate.
• Release matrix split into fast (native, Windows, cross — ~30 min) and slow (emulated/VM platforms — up to 75 min) waves
• New .cargo/config.toml: -lstdc++ for illumos (unrar_sys via cleave compiles C++ and rustc links -nodefaultlibs), -lgcc for 32-bit ARM musl (jemalloc's __ffsdi2 / byte-width CAS intrinsics).

Dependencies

• cleave 2.7.1 → 2.7.2, atomdrift-scan 2.7.0 → 2.7.1, plus filefacts, fletch, stng 1.8.0 → 1.9.0, libredox 0.1.19 → 0.1.20.

v0.4.0

Three new detectors, none rule-dependent

• Same-version binary repacks where the metrics move but the version doesn't.
• Encoded single-line payloads newly wired into a package entrypoint.
• Readable implants grafted in, caught by timestamp outliers rather than obfuscation.

Sharper judgment

• Reads versions from artifact metadata; without one, every release-pressure check was silently off.
• --deps findings now change the verdict instead of just printing.
• Dependency ranges resolve to current, not their floor — pinning to the floor is how you miss event-stream.

Fixes

• isomer ci --deps silently skipped both dependency folding and LLM interpretation.
• --offline didn't actually suppress the LLM.
• An unreadable base file could fabricate the signal that lowers a verdict.
• CI hardening: a caller-supplied --base could forge GitHub workflow outputs.
• Rules now auto-update like scan does.

Output

• Versioned features block in JSON for ML consumers.
• Trait grid ranked by criticality × confidence, capped at 24.
• Terminal row caps no longer truncate the data the verdict is computed from.

v0.3.0

Detection

• Shape signals: a dense capability jump in few files, or a package that deletes its own tree, now escalates alone.
• A patch or same-version repack that gains behavior is raised to High. Disproportion is the signal.
• Source archives read whole — build macros, fixtures, payload carriers stay in the diff.
• Two new shapes named: a build macro growing a joined shell eval, and an unchanged loader over a changed compressed carrier. The xz shape.
• Archive roots normalized, so pkg-1.2.3/ and pkg-1.2.4/ pair as one member.
• Cleanups score as cleanups: disabled handlers, removed traits, falling risk read Notable, not hostile.

False positives, gone

• absent → present identity is not drift. Only stripped identity is.
• Filename-only identity never compared against manifest identity.
• Empty identity objects likewise.
• Absolute ML risk no longer condemns. Only a worsening band moves the verdict.

The model

• Its verdict now raises displayed severity, and the risk bar with it. It never lowers.
• The gate stays deterministic. A hallucination must not fail your build.
• Raised numbers are labeled azoth+llm.
• Prompt rewritten around an evidence order, and taught what remediation looks like.
• --format interpret prints the exact payload and never calls out.

Output

• --explain removed. One view. If isomer speaks, it shows everything.
• Gained capabilities render as a tree: + new class, ↑ grown.
• Removed high-risk behavior gets its own section.
• Hunks name the member, dim the path, wrap instead of truncate.

Distribution

• 20 targets: Linux gnu/musl (x86_64 → riscv64, s390x, ppc64le, loongarch64, arm), macOS, Windows, the BSDs, illumos, Solaris, Haiku, Hurd.
• macOS binaries signed and notarized; a preflight proves the credentials before any build.
• Sigstore provenance attestation, checksums, per-target retry, dry runs.
• Runtime smoke tests on OpenIndiana, Tribblix, GhostBSD, openEuler, BlissOS.

Build

• New release-lto profile — fat LTO, one codegen unit — for shipped binaries. Everyday cargo build --release gets thin LTO.
• cleave and atomdrift-scan rev pins dropped; Cargo.lock holds the commit.

Full Changelog: v0.2.0...v0.3.0

v0.2.0

Added

• Rename-aware artifact pairing
• git rename detection in ci
• Scope reporting (source only / source + build output) so a run whose base build failed can't read as a clean full comparison.
• make validate-samples

Removed

• .isomer.toml suppression policy; unnecessary for differential analysis

Changed

• Clean CI runs emit a one-line step summary instead of a full report.

isomer 0.1.0

First release!