Harden Bash tool: preflight sanitizer for Windows portability and inline-source guards, plus output truncation

[mmprotest]

Motivation

• Commands emitted by agents were repeatedly failing on Windows because bash execution used cmd.exe semantics and Unix utilities or PowerShell cmdlets were being mixed incorrectly.
• Large multi-line or long python -c and here-string style commands caused brittle quoting/terminator failures and long feedback thrash.
• The model relied on shell-side pagination like | head to limit output, which should be handled by the runner instead of by agent-inserted Unix toolchains.

Description

• Added a surgical preflight sanitizer _preflight_shell_command in villani_code/tools.py that rejects brittle inline source-writing patterns (long or multi-line python -c, here-strings, large inline redirects) with an actionable recovery hint.
• Implemented Windows portability hygiene to reject Unix-only patterns (e.g. | head, | tail, grep, sed, heredocs, cat >, rm -rf, touch, pwd) when running under a win* platform, and detect PowerShell cmdlets used without an explicit PowerShell invocation.
• Added runner-side output truncation (MODEL_OUTPUT_CHAR_LIMIT) returning a truncated model-facing stdout/stderr with markers like [stdout truncated to N chars] while preserving full outputs in debug callback payloads and setting truncated appropriately.
• Improved tool feedback to return normal tool error results (not crashes) with a short, actionable message recommending use of Write/Patch for file edits and simpler Bash commands; updated tests in tests/test_tools_hardening.py to cover the new guards.

Testing

• Ran pytest -q tests/test_tools_hardening.py and all tests passed (8 passed) validating sanitizer behavior and truncation.
• Ran pytest -q tests/test_tools_schema.py and it passed (1 passed).
• Files changed: villani_code/tools.py and tests/test_tools_hardening.py.

---

Codex Task