fix(static_checks): Playwright C# assertions + real Windows CI proofs + C# test_file resolution (follow-up to #4)

[rlabarca]

Follow-up to #4, addressing the three review notes left on that PR. Stacked on fix/windows-static-checks-3 so the diff is just the new work; it will retarget to main automatically when #4 merges.

What this addresses

1. C# Playwright assertions false-flagged HOLLOW (review note)
check_csharp's no_assertions guard now recognizes Playwright fluent assertions — Expect(...) / Assertions.Expect(...) chained to a To*Async() matcher (ToBeVisibleAsync, ToHaveTextAsync, ToContainTextAsync, …). A bare Expect(x) with no matcher chain is still flagged. (static_checks RULE-31, PROOF-55.)
To keep the broader structural acceptance from masking weak tests, Pass-2 audit criteria gained a presence/visibility-only assertion WEAK check (audit_criteria.md → Criteria-Version 17).

2. Windows behavior was simulation-only (the original ask)
RULE-29/30 are now verified for real on a windows-latest GitHub Action (.github/workflows/windows-proofs.yml) via a new @windows proof tier — PROOF-53 (native msvcrt lock path, _HAS_FCNTL genuinely False) and PROOF-54 (--load-criteria under the native console codec) — committed back by CI, alongside the existing host simulations (PROOF-50/52). The schema_proof_format anchor (RULE-4) and spec_format.md (Format-Version 8) admit the platform-gated tier; the feature-scoped overwrite is per (feature, tier), so host runs that skip these tests never clobber the CI-emitted file.

3. Empty test_file blocked purlin:audit on C# (review note)
Under dotnet test, TestCase.CodeFilePath is often null, so the xUnit logger emits a blank test_file. 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); the audit skill uses it as a fallback. Native population still requires RunConfiguration.CollectSourceInformation=true + full PDBs — now documented. (static_checks RULE-32 / PROOF-56, skill_audit RULE-16 / PROOF-16.)

Also: bumped .claude-plugin/plugin.json to 0.9.5 to match VERSION (the manifest wasn't bumped during the v0.9.5 release; purlin_version caught the drift). Normalized the pytest proof plugin's test_file to forward slashes (no-op on POSIX).

Verification

• purlin:status: 40/40 features VERIFIED, integrity 100%.
• Windows Action: green on windows-latest — PROOF-53/54 pass natively.
• Independent audit: PROOF-53/54/55/56 STRONG; 0 WEAK/0 HOLLOW.
• No version bump (stays 0.9.5); the v0.9.5 tag is moved to the validated commit.

🤖 Generated with Claude Code

[LCbarsenault]

Validated v0.9.5 (branch tip 04cdca7) end-to-end against a real Blazor Server / xUnit / Playwright project — .NET 7, Windows 11, Python 3.12. The suite is 92 proofs across 14 features, every one with test_file: "", so C# auto-discovery gets exercised on every single proof. Summary: #1 and #3 fully fixed; #2 common path fixed (unblocks us); 3 resolver edge cases worth a look before merge.

1. Playwright assertions — FIXED ✅

• Nav_ShowsUser_AndSignOut (verifies only via Assertions.Expect(...).ToBeVisibleAsync(), no Assert.*) → pass / "structural checks passed". Same for the other Expect-only e2e methods. No more no_assertions/HOLLOW false flag.
• Negative control (added a throwaway fixture): a genuinely assertion-free method is still flagged — no_assertions: "test method has no Assert./.Should()/.Verify()/Expect(...).To*Async() call" — while a Playwright-only sibling passes. Correct in both directions.

3. Windows toolchain — CLEAN ✅

Every subcommand ran with no fcntl ModuleNotFoundError and no cp1252 decode/encode crash on Windows 11 / Python 3.12: --load-criteria (Criteria-Version 17), --read-cache, --check-proof-file, full Pass-1, and --resolve-source. The full purlin:audit ran end-to-end across all 14 features / 92 proofs (Pass 0.5 → Pass 1 → Pass 2) with no toolchain errors, 100% test_file auto-resolution, and 0 HOLLOW from Pass 1.

2. C# auto-discovery — common path FIXED ✅, edge cases below

The original blocker is gone: I resolved all 92 empty-test_file proofs via --resolve-source; 0 unresolved, 0 Pass-1 false HOLLOW. This includes the class-name≠file-name case (ReqExplorerCsvTests lives in ReqExplorerViewTests.cs).

Per your request I pushed on the edge cases with throwaway fixtures. resolve_test_file_from_name takes parts[-2] as the type and searches for \b(?:class|struct|record|interface)\s+<type>\b, picking the stem-match else the alphabetically-first file. That produces three mis-resolutions:

Shape	FQN tested	Resolved to	Result
Nested	Ns.Outer+Inner.M (xUnit's actual FQN)	—	❌ UNRESOLVED
Nested	Ns.Outer.Inner.M (dotted)	probe_nested.cs	✅
Generic	Ns.Gen`1.M (xUnit's actual FQN)	—	❌ UNRESOLVED
Generic	Ns.Gen.M (bare)	probe_generic.cs	✅
Partial (method in file _b)	Ns.P.MethodInB	probe_partial_**a**.cs	❌ wrong half
Same name, 2 namespaces	Probe.Alpha.Dup.M	dup_alpha.cs	✅ (only by alpha sort)
Same name, 2 namespaces	Probe.Beta.Dup.M	dup_**alpha**.cs	❌ namespace ignored

Root causes + suggested fixes:

• Same name across namespaces (highest impact). parts[-2] is the bare type name; the namespace is discarded, so Beta.Dup.M lands on the Alpha file. Duplicate simple names (Tests, Helpers, Handler) across namespaces are common. Fix: when >1 file declares the type, disambiguate with the namespace from the FQN — prefer a file whose content also has namespace ...Beta (or the namespace segment(s) before the type).
• Nested +. xUnit's fully-qualified name uses Outer+Inner; re.escape makes + literal so class Outer+Inner never matches. Fix: treat + (and .) as nested separators and search on the innermost type name.
• Generic `N. Backtick-arity suffix never matches class Gen<T>. Fix: strip a trailing `\d+ from the type token before building the regex. (Low priority — open generic test classes aren't runnable in xUnit, so this is mostly academic.)
• Partial split across files. Both halves declare class P; the resolver returns the alphabetically-first file, which may not hold the method, so Pass-1 then scans the wrong file. Fix: among type matches, prefer the file that also contains the method name; if still ambiguous, return all candidates.

These only bite projects that actually have nested/generic/duplicate-name/partial test classes — our committed suite has none, which is why the 92-proof sweep was clean. They're not blockers for us; flagging them as the pre-merge feedback you asked for.

Integrity held at 14/14 features VERIFIED, 100% throughout.

---

Install note (Claude Code side, not purlin): on a corp network that blocks SSH (port 22), the documented git@github.com: URL can't clone. Worked around it with git config --global url."https://github.com/".insteadOf "git@github.com:". Separately, /plugin marketplace add <url>#branch --scope project folded --scope project into the branch ref (fatal: Remote branch '<branch> --scope project' not found) — putting the flag first, or adding via a local clone path, avoided it. Both are CC-side; mentioning in case other testers hit them.

[rlabarca]

Confirmed working on the reporter's real .NET 7 / Blazor / xUnit / Playwright project, 40/40 features VERIFIED at 100% integrity, Windows Action green. Merging into main (right after #4) for v0.9.5.