Persistent ClaudeSession ignores the sub-skill system prompt, causing 'setup produced no tasks'

[FidoCanCode]

Regression introduced in #456. claude_start/claude_run in kennel/worker.py call build_prompt(fido_dir, subskill, context) which writes two files:

• fido_dir/system — persona.md + sub-skill.md (e.g. setup.md, task.md)
• fido_dir/prompt — the context (request, repo, branch, etc.)

In the pre-session code path, print_prompt_from_file spawns a new claude with BOTH files — the sub-skill instructions are loaded as system prompt.

In the persistent-session path (if session is not None), only the prompt file is sent:

prompt = (fido_dir / "prompt").read_text()
with session:
    session.send(prompt)
    session.consume_until_result()

The persistent ClaudeSession was constructed with only sub/persona.md as its system prompt, so claude never sees setup.md. Result: claude responds to a bare context with no instructions and doesn't call kennel task add. Then:

RuntimeError: setup produced no tasks

Watchdog restarts the thread, it tries again, same failure. Observed live after #475 landed, leaving fido stuck on #465 with 3+ abandoned branches.

Fix: prepend fido_dir/system content to the user message when routing through the persistent session, so claude gets the sub-skill instructions as part of each turn. Something like:

system_content = (fido_dir / "system").read_text()
prompt = (fido_dir / "prompt").read_text()
with session:
    session.send(f"{system_content}\n\n---\n\n{prompt}")
    session.consume_until_result()

Apply to claude_start and claude_run both. Test that setup, task, ci, resume, and comments sub-skills all deliver their instructions through the persistent session.