fix(copilot): detect truncated write_workspace_file and guide LLM to source_path

When the LLM tries to inline a very large file as content in
write_workspace_file, the SDK silently truncates the tool call
arguments to {}. The tool then returns a generic 'filename required'
error, which the LLM doesn't understand and retries the same way
(wasting 500s+ per attempt, as seen in session c465eff9).

Now: when ALL parameters are missing (likely truncation), return an
actionable error explaining what happened and how to fix it — write
the file to disk with bash_exec first, then use source_path to copy
it to workspace. This gives the LLM a clear recovery path instead
of a retry loop.