Add tool-call logging with verbose mode

[pnilan]

What

Adds structured logging for every PydanticAI tool invocation, giving operators
visibility into which tools the agent calls during a conversation. A new
-v/--verbose CLI flag enables debug-level output that includes tool arguments.

How

• patchwork/logging_config.py — New module with setup_logging(verbose) that
  configures a patchwork logger writing to stderr. Idempotent (safe to call
  multiple times without adding duplicate handlers).

• patchwork/cli.py — _make_event_handler() returns an async callback matching
  PydanticAI's EventStreamHandler signature. It filters for FunctionToolCallEvent,
  logs the tool name at INFO, optionally logs args at DEBUG in verbose mode, and prints
  a dim Rich indicator (⚙ tool_name) to the console. The handler is wired into
  agent.run_stream() via the event_stream_handler parameter. An argparse parser
  adds -v/--verbose, and errors now go through logger.exception().

• Tests — test_logging_config.py (6 tests) covers logger creation, handler
  levels, and idempotency. test_tool_logging.py (4 tests) covers tool name logging,
  verbose args, non-tool event filtering, and multiple events.

Recommended Review Order

1. patchwork/logging_config.py — small standalone module
2. patchwork/cli.py — event handler + CLI wiring
3. tests/test_logging_config.py
4. tests/test_tool_logging.py

[pnilan]

Two issues stood out in this patch:

1. --verbose is not actually the gate for user-visible tool logging. In patchwork/cli.py, console.print(f"[dim]⚙ {tool_name}[/dim]") runs for every FunctionToolCallEvent, even when verbose=False. That changes the default CLI transcript for all users, not just users who opted into verbose mode, and it will interleave extra lines into any consumer expecting only the assistant stream. If the goal is to show tool activity only in verbose mode, I'd gate that print behind verbose or introduce a separate explicit flag like --show-tools.

2. setup_logging() can't honor a later verbose=True call once the logger has already been initialized. patchwork/logging_config.py returns early whenever handlers already exist, so the handler level from the first call is frozen. A repro is setup_logging(False) followed by setup_logging(True): debug tool-arg logs remain disabled even though the second call asked for verbose mode. I'd keep the idempotent single-handler behavior, but always update the handler level/formatter on subsequent calls instead of returning immediately.

I'd also add one regression test for each behavior:

• a test proving non-verbose mode does not print tool indicators
• a test proving setup_logging(False) followed by setup_logging(True) enables debug output on the same logger

If the always-on tool indicator is intentional, I think that should be called out explicitly in the PR description because it's a UX change, not just logging infrastructure.