Adopt rtk-ai/rtk#3029: runner - surface failing tool's stderr under stdout-only filters

[kylehgc]

Closes #6.

Adopts rtk-ai/rtk#3029 (cherry-pick, authorship preserved) plus one fork-authored amendment.

Commits

• c27bbd0 fix(runner): surface a failing tool's stderr under stdout-only filtering — cherry-pick of upstream 328bb53, author georgyia (@georgyia). Callers pairing stdout_only with early_exit_on_failure (ls, tree) already printed stderr on failure; bare stdout_only callers (pytest, ruff, prettier, go test, golangci-lint, phpunit, phpstan, pint, rubocop) captured it and dropped it. The fix emits raw_stderr when the child fails; success stays suppressed. + tests/stderr_passthrough_test.rs (cfg(unix), shim-based, failing-before).
• afc5258 test: use expect with reason instead of bare unwrap in stderr test — amendment: .claude/rules/rust-patterns.md requires expect("reason") over bare unwrap() in tests; mechanical rename of the five call sites in the new test file.

Adoption checklist

• Sync: no-op — upstream/develop is already an ancestor of fork develop
• Repro on pre-cherry-pick binary (below)
• Cherry-pick preserving authorship; fixed behavior confirmed with real output (below)
• Quality gate: cargo fmt --all clean, cargo clippy --all-targets zero warnings, cargo test --all 2466 passed / 0 failed
• Amendment as separate fork-authored commit

Repro (real output, this Windows machine)

This machine has python but no pytest — the exact rtk-ai#3026 quadrant, no uv needed: rtk pytest falls back to python -m pytest, the spawn succeeds, and the child's diagnostic goes to the captured-and-dropped stderr.

Before (develop @ 1018b8b):

$ rtk pytest --version
exit 1, stdout 1 byte (a single newline), stderr 0 bytes

A completely mute failure — reads as a quiet pass.

After (this branch):

$ rtk pytest --version
exit 1, stderr 48 bytes:
C:\Python314\python.exe: No module named pytest

Review

Two-axis review (standards + spec) run pre-PR.

• Spec: clean — fix hunk byte-identical to upstream and correctly placed (early-return ordering verified), no-double-print claim verified against FilterMode::CaptureOnly and both raw_stderr print sites, both claimed tests present and asserting what they claim, no scope creep (git diff pr-3029 HEAD empty on both files).
• Standards: one hard finding (bare unwrap() in tests vs rust-patterns.md) fixed by the amendment. Known ceilings, noted not fixed: the new tests are #![cfg(unix)] so Windows CI never runs them (the runner change is platform-neutral and the Windows behavior was verified manually above); emitted stderr is not counted in rtk gain tracking on failure (consistent with the pre-existing skip_filter_on_failure branch).

🤖 Generated with Claude Code