Validate quoted shell test commands without false path rejections

[wwind123]

Summary

Validate explicit quoted shell command strings as commands, rather than treating the whole string as one filesystem path.

Reproduction

The historical llm-dialectic #1136 / PR #1137 run pushed its implementation and then failed test-report validation on:

bash -lc '/home/wwind123/llm-dialectic/.venv/bin/pytest tests/test_grafana_dashboards.py tests/test_production_topology_artifacts.py -q'

Reproduced against current main 4b46054, importing the guard explicitly from this checkout with PYTHONPATH=src. The error names the entire quoted command, including test arguments, as an outside-checkout path. The external virtualenv executable should retain the existing interpreter exemption, while test paths and working-directory changes must still be checked independently.

Related to #493. This does not close that broader environment-prose issue or change #762's chained managed-wrapper contract.

Changes

• Recognize explicit sh, bash, and zsh command strings using -c and a conservative set of no-operand option combinations such as -lc.
• Split and validate the nested executable text through the existing clause/path/URL checks; retain checks on the launcher, prefix environment, and positional arguments.
• Preserve rejection of outside test paths, outside working directories, live URLs, and unknown outside scripts. Unsupported shell option forms receive no special handling.
• Bound nested shell expansion to eight levels and reject malformed command-string quoting.
• Cover the historical example, wrapper prefixes, nested shells/managed wrappers, trailing clauses, and negative cases. Document the behavior and limits.

This validates reported command text. It neither executes shell strings nor evaluates substitutions nor supplies a shell sandbox. No live service tests or provider calls are involved.

Validation

• Focused guard/runtime/rendering suites: 365 passed using PYTHONPATH=src and isolated writable runtime-lock storage.
• Full suite: 2,898 passed in 146.61 seconds with env PYTHONPATH=src XDG_RUNTIME_DIR=/tmp/quoted-shell-test-runtime python3 -m pytest tests/ -q --tb=short. Existing tests require user-cache log writes, so the full suite ran with that permission.
• git diff --check: passed.
• An initial focused run had four runtime-test failures because the sandbox could not write lock files under /run/user/1001. With XDG_RUNTIME_DIR=/tmp/quoted-shell-test-runtime, all focused tests passed; no live locks were removed or runtime-lock implementation changed.
• An initial new test mistakenly created the same fixture directory twice; corrected the test to reuse the fixture path.

-- OpenAI Codex