fix(dotnet): treat global.json MTP as VsTestBridge instead of native mode

[danielmarbach]

Fixes #1289

Summary

Problem

PR #746 introduced MtpNative mode which injected --report-trx as a direct flag for repos with global.json MTP mode ("test": { "runner": "Microsoft.Testing.Platform" }). This broke xUnit v3 projects with exit code 5:

Unknown option '--report-trx'
Test run completed with non-success exit code: 5

The Altinn/altinn-register repo was affected: .NET 9 SDK, global.json MTP mode, xUnit v3 — no project-file MTP properties at all.

Root cause

--report-trx is not a universal MTP flag. It requires the Microsoft.Testing.Extensions.TrxReport NuGet package which xUnit v3 doesn't bundle (it has --report-xunit-trx instead). MSTest and TUnit bundle it, so it worked for them but failed for xUnit v3.

Verified behavior

All tested with global.json having "test": { "runner": "Microsoft.Testing.Platform" }:

Framework	SDK	-- --report-trx	Direct --report-trx
xUnit v3 3.2.2	.NET 9	✗ Unknown option	✗ Unknown option
xUnit v3 3.2.2	.NET 10	✗ Unknown option	✗ Unknown option
MSTest 4.0.2	.NET 9	✓	N/A (MSBuild error)
TUnit 1.36.0	.NET 9	✓ TRX generated	N/A
TUnit 1.36.0	.NET 10	✓ TRX generated	✓ TRX generated

The xUnit v3 issue reporter confirmed (xunit/xunit#3456) that adding Microsoft.Testing.Extensions.TrxReport NuGet makes --report-trx work universally.

Additionally, -nologo breaks on .NET 10 native MTP with "Unknown option '--nologo'", so it's now skipped for all MTP mode.

Fix

Simplified TestRunnerMode from three variants to two:

• Classic: VSTest runner → --logger trx --results-directory, -nologo injected
• MtpVsTestBridge: Any MTP detection (global.json OR project-file) → -- --report-trx, -nologo skipped

Removed MtpNative and MtpGlobalJson. Both global.json MTP and project-file MTP properties now use MtpVsTestBridge, injecting -- --report-trx after the separator.

For xUnit v3 without Microsoft.Testing.Extensions.TrxReport, this produces an informative "Unknown option" error — the user knows to add the package. This is preferable to silently skipping TRX injection, which would regress token savings for the majority of MTP users (MSTest, TUnit) who do bundle TrxReport and benefit from the optimized TRX parsing path. Detecting the NuGet package from csproj/Directory.Build.props/Directory.Packages.props was ruled out as impractical (MSBuild inheritance, central package management).

Test plan

• cargo fmt --all && cargo clippy --all-targets && cargo test
• Manual testing: rtk <command> output inspected

Important: All PRs must target the develop branch (not master).
See CONTRIBUTING.md for details.

[pszymkowiak]

[w] wshm · Automated triage by AI

📊 Automated PR Analysis

🐛 Type	bug-fix
🟡 Risk	medium

Summary

Fixes global.json MTP detection to use VsTestBridge mode instead of the removed MtpNative mode, which was injecting --report-trx directly and breaking xUnit v3 projects with 'Unknown option' errors. Also skips -nologo for all MTP modes since it breaks on .NET 10 native MTP, and refactors the project file scanning into a separate function.

Review Checklist

• Tests present
• Breaking change
• Docs updated

Linked issues: #1289

---

Analyzed automatically by wshm · This is an automated analysis, not a human review.

[F0rty-Tw0]

The compatibility matrix in the description made me back off two concerns I'd flagged on first read:

1. -nologo skip across MTP modes: I worried this was over-conservative, until the .NET 10 native MTP "Unknown option --nologo" repro. Skip is correct.

2. xUnit v3 "Unknown option" surfacing: initially read as a silent UX downgrade, but the trade-off is conscious. A silent skip would regress token savings for the MSTest/TUnit majority who do bundle TrxReport, and the error is informative once a user knows the package name. NuGet detection from csproj/Directory.Build.props/Directory.Packages.props is impractical for the cited reasons (MSBuild inheritance, central package management), so surfacing the error is the right call.

I also ran a fresh sub-review against .claude/rules/rust-patterns.md and cli-testing.md. Variant deletion is complete: every MtpNative site the diff touches is a removal or rewrite, no residue. The scan_project_files_for_mtp early-return refactor is behavior-preserving (no multi-project regression), and there are no new unwrap() outside tests.

Three small follow-ups follow as inline comments, none blocking.

Open since April 18 with P2-important, three-OS CI green, repro from Altinn/altinn-register. Would be great to see this land.

[F0rty-Tw0]

After this PR, MtpVsTestBridge covers any MTP detection: TestingPlatformDotnetTestSupport=true, UseMicrosoftTestingPlatformRunner=true, UseTestingPlatformRunner=true, and global.json MTP. Per the PR description matrix, several of these don't go through an actual VSTest bridge for MTP at all (the previous code classified some as MtpNative).

The variant doc handles the dual coverage well, but the name biases readers toward the bridge interpretation when the variant is really describing an injection strategy (-- --report-trx), not a single runtime.

Two options to avoid confusing a future reader:

• Rename to MtpAfterSeparator (descriptive of the action: inject -- --report-trx) rather than the runtime.
• Or add a // Naming note: line stating that "VsTestBridge" here is a historical name kept for diff minimization, covering both bridge and native MTP cases.

Cheap, prevents future churn.

[F0rty-Tw0]

The three new tests exercise helpers (parse_global_json_mtp_mode, scan_mtp_kind_in_file, scan_project_files_for_mtp) but none drives the full detect_test_runner_mode returning MtpVsTestBridge from the global.json-only path that motivated the PR.

is_global_json_mtp_mode walks the current directory via std::env::current_dir(), which makes an end-to-end test awkward without test isolation (set_current_dir + serial_test). Every other detection branch (csproj, Directory.Build.props) has an end-to-end test, and global.json is the actual regression path here.

Cleanest fix: refactor is_global_json_mtp_mode() to take a starting-dir parameter and have a thin wrapper read current_dir(). That unblocks unit testing without filesystem juggling and lets a future test exercise the full priority chain (global.json → MtpVsTestBridge) end-to-end.