v0.9.5 — Cross-platform audit Pass-1 (Windows) & C# support

Fixed

• purlin:audit now runs on Windows. The audit toolchain shells out to scripts/audit/static_checks.py, which was unusable on Windows for three independent reasons, each fixed here (issue #3):
  • import fcntl crashed the module on load. fcntl is POSIX-only, so every subcommand (--read-cache, --load-criteria, --check-proof-file, Pass 1, --write-cache, --prune-cache) died with ModuleNotFoundError before doing any work. fcntl is now imported under try/except ImportError behind a _HAS_FCNTL flag, and the audit-cache lock goes through _lock_exclusive/_unlock helpers that use fcntl.flock on POSIX and msvcrt.locking on Windows — preserving the exclusive read→merge→write serialization that keeps concurrent subagent writers from clobbering each other (static_checks RULE-25/RULE-29).
  • Text I/O assumed UTF-8. ~16 open() calls omitted encoding=, so on a Windows cp1252/ASCII locale they raised UnicodeDecodeError reading the tool's own UTF-8 files (e.g. references/audit_criteria.md). Every text-mode open() now specifies encoding='utf-8' (static_checks RULE-30).
  • --load-criteria printed non-ASCII to a non-UTF-8 console. Even after the read fix, the criteria text (which contains ✓/⚠ glyphs) raised UnicodeEncodeError on an ASCII stdout. static_checks.py now reconfigures stdout/stderr to UTF-8 at startup (static_checks RULE-30).

Added

• Deterministic Pass-1 coverage for C#/.NET tests. Pass 1 previously recognized only pytest/Jest/Shell markers, so xUnit/NUnit/MSTest tests carrying [Trait("PurlinProof", "feature:PROOF-N:RULE-N:tier")] produced zero proofs and the structural checker silently no-op'd on .cs files. The new check_csharp analyzer parses those trait markers, locates each [Fact]/[Theory] method body via a C#-aware brace/string scanner, and runs assert-true / no-assertion detection — recognizing xUnit Assert.*, NUnit Assert.That, MSTest Assert.*, FluentAssertions .Should(), and Playwright fluent assertions (Expect(...)/Assertions.Expect(...) chained to a To*Async() matcher) as assertions (static_checks RULE-31). A bare Expect(x) with no matcher chain is still flagged no_assertions. The dispatch was extracted into analyze_test_file() and now routes .cs. references/supported_frameworks.md documents the new coverage. To keep this broader structural acceptance from masking weak tests, the Pass-2 audit criteria gained a "presence/visibility-only assertion" WEAK check (references/audit_criteria.md → Criteria-Version 17).
• Portable interpreter guidance in purlin:audit. skills/audit/SKILL.md documents a python3 → python → py -3 fallback for invoking static_checks.py, since stock Windows does not put python3 on PATH (skill_audit RULE-15).
• purlin:audit reaches C# tests even when test_file is empty. Under dotnet test, TestCase.CodeFilePath is often null (no surfaced source info), so the xUnit logger emits proof entries with a blank test_file — which previously blocked the audit skill from locating the .cs code to scan (Pass 1) or read (Pass 2). static_checks.py now exposes --resolve-source <test_name>, which derives the declaring type from the fully-qualified test_name and locates its .cs file (skipping bin/obj), and skills/audit/SKILL.md uses it as a fallback (static_checks RULE-32, skill_audit RULE-16). Populating test_file natively still requires RunConfiguration.CollectSourceInformation=true with full PDBs — now documented.

Testing

• New proofs: static_checks PROOF-44/45 (AST structural guards — no unconditional fcntl import; encoding='utf-8' on every text-mode open()), PROOF-46–49 (C# assert-true / no-assertion / multi-framework assertion recognition / .cs dispatch), PROOF-50 (Windows lock path driven through an injected fake msvcrt, @integration), PROOF-51 (full CLI pipeline via real subprocess, @e2e), PROOF-52 (the cp1252 failure reproduced by running --load-criteria under PYTHONUTF8=0/LC_ALL=C, @e2e); skill_audit PROOF-15. PROOF-55 covers the Playwright assertion recognition (and the bare-Expect negative). Windows behavior is now verified for real on a windows-latest GitHub Actions runner — .github/workflows/windows-proofs.yml runs PROOF-53 (native msvcrt lock path, _HAS_FCNTL genuinely False) and PROOF-54 (--load-criteria under the native console codec) in the dedicated @windows proof tier and commits the results back — in addition to the host simulation (fake msvcrt) and ASCII-locale subprocess repro. Independent audit of the new proofs: STRONG at 100% integrity, 0 WEAK/HOLLOW.

v0.9.4 — Plugin-bundled MCP server & e2e proof quality

Fixed

• MCP server no longer version-pinned in consumer projects. purlin:init wrote a project-level .mcp.json entry with an absolute path into the versioned plugin cache (~/.claude/plugins/cache/purlin/purlin/<version>/scripts/mcp/purlin_server.py). Project-scope .mcp.json takes precedence over plugin-provided servers and old cache directories stick around, so every plugin update silently stranded the project on the previous release's server — fresh data, stale code, no error (this is exactly how a 0.9.3 project kept rendering dashboards with 0.9.1 logic). The server is now bundled with the plugin itself: .claude-plugin/plugin.json declares it under mcpServers with ${CLAUDE_PLUGIN_ROOT}, which Claude Code resolves to the installed plugin path on every launch, so the server always tracks the installed version (skill_init RULE-38/39). Init no longer writes a purlin entry into the project's .mcp.json; it removes the legacy entry while preserving any other MCP servers in the file (skill_init RULE-40).

  Upgrading an existing project: after updating the plugin, run purlin:init --mcp once (new flag — runs only the migration step) to remove the stale entry, then /reload-plugins (or restart the session).

• sync_status flags the legacy entry automatically. When the project's .mcp.json defines a purlin server pinned to a plugin-cache path, every status report opens with an advisory naming the pinned path and a → Run: purlin:init --mcp directive — so pre-0.9.4 projects surface the migration on their first status call instead of relying on release notes (sync_status RULE-38). Dev checkouts (purlin entries pointing outside the plugin cache) are exempt.

Added

• E2E proofs must be observable flows. Reverse-engineering UI-heavy codebases with purlin:spec-from-code produced implementation-coupled proofs tagged @e2e ("Assert config.ts AES-decrypts...", "Assert loginRedirect uses scope X") — satisfiable without launching the app, steering builders into unit-style tests that audit WEAK/HOLLOW. The quality guide now has a canonical "E2E proof descriptions (observable flows)" section (references/spec_quality_guide.md): @e2e proofs read as arrange → act → observe through the real running app, never name source files or internal functions, observe data contracts at the boundary they cross (outbound request, rendered output, storage state after a real flow), and stay tool-agnostic — executable by Playwright, Cypress, an MCP-driven browser, or screenshot + vision (purlin_references RULE-13).
• E2E Proof Tier Integrity audit criteria. references/audit_criteria.md (Criteria-Version 16) adds two Pass 2 WEAK criteria at HIGH priority applying to ALL @e2e proofs, not just design anchors: tier mismatch (test tagged @e2e never drives a browser/renderer/full stack — or the converse, a render/flow test with no tier tag) and source-constant assertion (asserting a config constant where the rule describes runtime behavior) (purlin_references RULE-14). The version bump does not invalidate audit caches — the cache key excludes the criteria version.
• E2E runner reality check in purlin:spec-from-code. Phase 1 records an e2e_capable flag; when a category's generated proofs include @e2e and no e2e runner is detected, the skill warns in the category review block and the Phase 4 summary instead of silently emitting unrunnable proofs (skill_spec_from_code RULE-32). references/supported_frameworks.md gains an "End-to-end (browser) proofs" section documenting interim proof emission through the existing Vitest/Jest markers or shell purlin_proof wrappers (purlin_references RULE-15).

Changed

• purlin:spec-from-code enforces proof quality at generation time. The step 7 tier review now runs an inverse check — every @e2e description must match its tag or be rewritten/retagged (skill_spec_from_code RULE-30) — and step 11 validation rejects proof descriptions naming source files or internal symbols (skill_spec_from_code RULE-31). purlin:spec applies the same observable-flow check in Validate-Before-Commit (skill_spec RULE-8).

Testing

• New proofs: skill_spec_from_code PROOF-42..45, skill_spec PROOF-9, purlin_references PROOF-13/14/15 (grep guards over the skill and reference text). Independent audit of the new proofs: 0 WEAK/HOLLOW (all structural documentation guards, excluded from integrity scoring); 3 advisory regex-precision findings applied before commit. Full suite 372 passed, 40/40 features VERIFIED.
• skill_init PROOF-40/41 now parse .claude-plugin/plugin.json directly (asserting mcpServers.purlin uses python3 with ${CLAUDE_PLUGIN_ROOT} args) and PROOF-42 verifies the legacy .mcp.json migration instructions, replacing the old greps for project-level .mcp.json creation.

v0.9.3 — Dashboard visibility before tests exist

Quality-of-life release fixing the "empty dashboard" experience after purlin:spec-from-code: a freshly specced project now shows its full rule set and coverage plan in purlin-report.html before a single test has run.

Added

• Planned proofs in the dashboard. The report data (.purlin/report-data.js) now includes the PROOF-N entries declared in each spec's ## Proof section that have no executed result yet — status "planned", tier parsed from the proof's @tag (default unit), empty test location (report_data RULE-8). The dashboard renders them greyed with a "not run" tag in the Proof column instead of a bare dash (purlin_report RULE-33), so the intended coverage plan is visible immediately after spec generation. Planned proofs are display-only: proved/total counts, vhash, rule status, and feature status are computed from executed proofs exactly as before (report_data RULE-22). Proof results are still written only when tests run (purlin:unit-test / purlin:build / purlin:verify).

Changed

• Category sections are expanded by default. Previously every category started collapsed, so a project with many small categories looked like it had no specs at all. Categories now render expanded; collapsing one is remembered per browser via localStorage (purlin_report RULE-20, persistence unchanged per RULE-21).
• purlin:spec-from-code no longer creates single-spec folders. A new Phase 2 taxonomy step merges single-feature categories into the closest related category, or — when nothing fits — places the spec directly at specs/<name>.md with no folder (skill_spec_from_code RULE-29). Uncategorized specs display under "other" in the dashboard.
• Docs and skill wording aligned with actual behavior. PASSING criteria say "all rules" (matching _determine_status); verify receipt format documented as features=N/T; installation/testing/lifecycle guides corrected (committed digest, real GitHub Actions example, shell plugin path); spec-from-code guide documents the post-generation build lifecycle; lifecycle PM diagram regenerated.

Testing

• New proofs: planned-proof emission and dedup against executed results including required/global anchor rules (report_data PROOF-22), coverage isolation (PROOF-23), greyed "not run" rendering with no audit tag (purlin_report PROOF-33, Playwright), and the single-spec folder ban (skill_spec_from_code PROOF-41). Dashboard PROOF-19/20/21 updated for the expanded-by-default behavior. Independent audit of the changed proofs: 5 STRONG, 0 WEAK/HOLLOW.

v0.9.2 — .NET test support (xUnit)

Incremental release adding .NET to the supported test ecosystems, backed by a refactor of the proof-plugin specs.

Added

• xUnit/.NET proof plugin. scripts/proof/xunit_purlin.cs is a custom dotnet test logger (ITestLoggerWithParameters, FriendlyName purlin) that collects proof traits in-process — no .trx post-parsing — and emits feature-scoped proof JSON per the shared proof-plugin contract. Mark tests with a trait:

  [Fact]
  [Trait("PurlinProof", "feature_name:PROOF-1:RULE-1:unit")]
  public void ValidLogin() { ... }

  Because the marker is a test trait rather than a parsed title string, NUnit [Category]/[Property] and MSTest [TestProperty] surface the same way — the logger covers C#, F#, and VB.NET test projects. Run with dotnet test --logger purlin -- RunConfiguration.CollectSourceInformation=true. Setup is manual for now (the .NET test platform only discovers loggers from assemblies named *TestLogger.dll) — see references/formats/proofs_format.md for wiring steps. Proven by an integration suite that drives a real dotnet test --logger purlin run; independent audit: 6/6 STRONG.

• purlin:init presents every shipped framework. The framework selection list now covers all shipped plugins — pytest, Jest, Vitest, C, PHP, SQL, Shell, and xUnit (plus "other") — instead of only the auto-detected subset (skill_init RULE-48). Frameworks that need manual wiring (xUnit) print their setup steps after the plugin file is copied.

Changed

• Proof-plugin specs split per framework. The monolithic proof_plugins spec is gone, replaced by one spec per plugin (proof_plugins_pytest, _jest, _vitest, _shell, _c, _php, _sql, _xunit) plus a proof_common anchor that holds the shared contract: feature-scoped overwrite, tier file naming, marker→JSON field mapping. Proof markers were re-homed to their per-plugin specs.
• references/supported_frameworks.md is now Format-Version 5: adds an Additional Plugins (manual setup) section for shipped-but-not-auto-scaffolded plugins (currently xUnit).

Testing

• New integration proofs drive the compiled xUnit logger through a real dotnet test run (dev/test_multilang_proof_plugins.py::TestXUnitProofPlugin).
• Strengthened existing proofs: C plugin proofs regenerated from a real gcc run, Vitest RULE-1 marker parse now asserted inline, pytest "call phase only" boundary proved, and three audit-pipeline proofs recorded from a full-suite run.

v0.9.1 — Vitest reporter & JS/TS audit fixes

Bug-fix release addressing two reported issues in JavaScript/TypeScript proof handling.

Fixed

• Vitest reporter silently emitted zero proofs on Vitest 2.x+ (#1). scripts/proof/vitest_purlin.ts collected proofs in onTaskUpdate, whose pack shape changed in Vitest 2 ([id, result, meta], where result lacks name/file), so the marker regex never matched and tests passed with no .proofs-*.json written. The reporter now collects in onFinished(files) via a recursive file→suite→test walk — the hook whose shape is stable across Vitest 2.x–4.x — mapping result.state to pass/fail, skipping unrun tasks, and resolving test_file from the file task's filepath.
• All Vitest projects now scaffold the native vitest_purlin.ts. Vitest never calls Jest's onTestResult/onRunComplete hooks, so the previous "Vitest → jest_purlin.js" mapping was also silently broken. Vitest loads .ts reporters natively via Vite, so one reporter covers both JS and TS projects.
• check_js audit regex misfired on common Vitest patterns (#2). scripts/audit/static_checks.py::check_js() used a lazy (.*?)\}\s*\) body matcher that truncated at the first inner } (options objects, destructured params, type assertions) and a [^"']* title class that dropped titles containing apostrophes. Replaced with a stdlib brace-balancing tokenizer that tracks string/template/regex literals and comments.
• Claude plugin manifest version lagged the release. .claude-plugin/plugin.json — the version consumers install against via the plugin marketplace — was left at 0.9.0. It now reads 0.9.1, and a new guard (purlin_version RULE-5 / PROOF-5) asserts the manifest version stays in lockstep with the VERSION file, joining the existing templates/config.json check so this can't silently recur.

Testing

• PROOF-29 was strengthened to actually drive the compiled reporter via a synthetic Vitest 2.x+ task tree (the old proof only ran tsc + node on hand-built JSON and never invoked the reporter). New regression proofs cover the exact issue #2 repro.
• Docs aligned: references/supported_frameworks.md (Format-Version 3), references/formats/proofs_format.md, and docs/testing-workflow-guide.md now describe the native Vitest reporter and tested version range.