counselos plugs a quiet leak in Mike's debug logs

Mike's chat tooling has long printed verbose debug lines as it reads documents, builds context, and streams answers from the underlying AI models. Useful when you're developing - quietly dangerous in production, where anything written to stdout tends to get swept up by log aggregators, SIEM systems, and backup tapes. For a fork like counselos/mike-inhouse, pointed at privileged client work, that's a paper trail nobody asked for.

The fix is small and disciplined: every log line that carries actual document content or storage paths is now gated behind a debug flag that's off by default. Read failures still surface so operators can tell when something breaks - they just no longer come with the document attached. A short note in the code spells out the threat model so the next contributor doesn't undo it by accident.