v0.1.0

Added

• Initial implementation (feat(init): initial implementation): the
  frontend-agnostic effect-analysis engine (engine/), the bash/POSIX and
  PowerShell frontends (front/), the embedded YAML knowledge base
  (kb/data/*.yaml), command resolution and flag/operand binding (bind/), the
  composition facade (internal/analysis/) and the flowsh CLI (cmd/flowsh/).
• The conformance corpus under testdata/corpus/ (GuardFall, destructive, benign
  and PowerShell cases) with a no-silent-miss regression gate.
• A tag-triggered release pipeline (.github/workflows/release.yml and
  .goreleaser.yaml): it validates the vX.Y.Z tag shape, re-runs go vet and
  the test gates on the tagged commit, cross-compiles linux/amd64, linux/arm64,
  darwin/arm64 and windows/amd64 with GoReleaser, emits a SHA-256
  checksums.txt and publishes a GitHub Release, followed by an SLSA
  build-provenance attestation verifiable with gh attestation verify. The
  build is reproducible (no injected date or commit hash, mod_timestamp pinned
  to the commit time) and stamps the tag into the binary through the additive
  main.buildVersion, so flowsh --version reports the release without
  affecting the frozen report-contract versions.

Changed

• Hardened the knowledge-base loader to tolerate CRLF/CR input and pinned
  repository line endings with a .gitattributes (* text=auto eol=lf).
• Enforced lint and dead-code gates (pinned .golangci.yml, a deadcode gate and
  documentation hygiene) and dropped the unreachable exported symbols they flag;
  the whole-program binder walkers and the corpus harness moved into test-only
  files and the new internal/corpus package, and the CLI's terminal output is
  routed through a single fprint/fprintln helper.
• Refined prose and punctuation across the contributor guide, README and user
  documentation.
• Refreshed the README effect-report examples and CLI usage: documented the
  --file flag in the synopsis and options table and updated the annotated
  examples (the root provenance line, the expanded score dimensions, the
  resolution/destructive lines and proven exfil: pairs; why,
  resolution and destructive fields in the JSON sample; the rm -rf $HOME
  score corrected to Critical).

Fixed

• Closed the silent-miss soundness gaps found in code review across the binder,
  both frontends, the engine and the knowledge base. The dominant class violated
  the no-silent-miss invariant: a command whose invocation matched no
  effect-contributing parameter produced neither an effect nor ⊤, so the report
  failed open to a benign result (bare reboot, kill -9 -1, xargs rm,
  builtin eval).
  • Binder. Matches a leading positional operand by value before falling
    back to index order, emits an intrinsic ProcSpawn for a resolved command
    that matched nothing, lowers kind:self parameters unconditionally, keeps a
    declared numeric flag as a flag rather than an operand, and derives the
    conservative flag from a ⊤ effect.
  • bash frontend. Confines control-flow signals to subshells, background
    jobs and speculative branches; inverts !; resets the status per statement;
    runs substitutions inside arithmetic/conditional/let and C-style for;
    executes array-element and subscript substitutions; unwraps builtin; treats
    >&word as a file redirect; and matches case patterns with a
    bash-compatible glob matcher.
  • PowerShell frontend. Resolves aliases in source order and
    case-insensitively, shadows a cmdlet by a declared alias or function,
    degrades comma lists and switch to ⊤, and models no-space redirections,
    data-file parameters and remote-execution cmdlets.
  • Engine. Made the effect key injective, kept Normalize's why-traces in
    step with the merged effects, and recognized private-key paths and Windows
    volume roots.
  • Knowledge base. Added an intrinsic self effect for the power-control
    commands, the long spellings and siblings of the destructive flags, and
    rejected the contradictory option+valueFrom:args combination.
• Made the PowerShell per-parse wall-clock budget advisory under the race
  detector (build-tagged constants disable the deadline check under -race
  while termination stays guaranteed by the parser's deterministic limits;
  ADR-0012).
• Widened the CI load-latency budget to 50 ms after measuring the hermetic,
  in-memory knowledge-base load on shared runners, superseding ADR-0004's 5 ms
  threshold while leaving the embed-and-version decision intact (ADR-0011).
• Skipped the sub-tick parse-budget assertions on hosts whose monotonic clock is
  too coarse to observe the deadline, and widened the repeated-statement source
  so the timeout-to-⊤ path stays covered everywhere (ADR-0013).