fix: prevent Windows subprocess output decode crashes

[xiarimr]

Problem

subprocess.run and subprocess.Popen decode text output with the platform's active encoding and strict error handling by default. On Windows, mixed or malformed command output can therefore raise UnicodeDecodeError; the original UTF-8-to-GBK fallback still raised for bytes invalid in both codecs and only covered s07.

Solution

• Keep subprocesses in text mode and use the platform's active encoding with errors="replace".
• Apply the policy consistently to all 42 text-output calls in the lessons, reference agents, and Agent Builder templates.
• Refresh the generated course source snapshot.
• Add regression coverage for UTF-8, GBK, invalid stdout/stderr bytes, empty output, timeouts, and future policy omissions.

Verification

• python -m pytest tests -q: 287 passed, 100 subtests passed
• Python 3.11 compile check: passed
• npm run build: passed
• Real API tool-use loop: invalid subprocess output became �, message roles remained valid, and the model completed normally

[Bill-Billion]

Thanks for identifying and reproducing the Windows decoding failure. I merged the latest main and extended the fix across all lesson and reference subprocess paths, with regression coverage and full Python, web-build, and real-model verification now passing.