feat(adapters): add Aider CLI adapter (--harness aider)

[Abhiix0]

Summary

Adds AiderAdapter, a new optional compatibility backend for the Aider CLI, following the same AgentAdapter contract as the existing Claude Code / GenericAgent / mini-swe-agent adapters. Aider has a large, active user base and no existing adapter, so this extends Bayesian Skill evolution to another widely-used coding harness without vendoring any of its code.

What's added

• bayesian_agent/adapters/aider.py — AiderAdapter dataclass:
  • Drives Aider non-interactively via --message-file (avoids ARG_MAX issues with large skill-context prompts and leaves an inspectable prompt file per workspace, instead of --message/argv).
  • Process-group-isolated subprocess execution with SIGTERM→SIGKILL timeout handling, matching ClaudeCodeAdapter's cleanup behavior (including killing orphaned child processes on timeout).
  • parse_result extracts token usage and cost from Aider's stdout (Tokens: X sent, Y received. Cost: $Z session.), including k/m suffix handling, since Aider's CLI output isn't JSON like Claude Code's.
  • load_run_from_workspace for resuming/repairing from prior runs.
  • Never imports the aider package — pure CLI subprocess boundary.
• Registered in bayesian_agent/adapters/__init__.py.
• Wired into experiments/run_benchmarks.py: new --harness aider choice, --aider-cli / --aider-timeout / --aider-auto-commits / --aider-yes-always flags, backend construction branch, and agent_name_for_harness entry.
• docs/adapters.md: new "Aider Adapter" section matching the existing adapter doc structure.
• README.md: Aider added alongside GenericAgent / mini-swe-agent / Claude Code in the feature list, compatibility-backends list, and file listing.
• tests/test_adapters.py: 7 new tests mirroring the existing adapter test conventions — command construction, token/cost parsing, task boundary, timeout handling, child-process cleanup on timeout, workspace-log recovery, and the "does not import eagerly" guarantee. All use a fake CLI script as cli_path, no real aider install required.

Testing

python3 -m pytest tests/ -q
89 passed

All 89 existing + new tests pass, including the 7 new Aider-specific tests and the full existing suite (no regressions in other adapters).

Notes

• CLI flags (--message-file, --yes-always, --auto-commits/--no-auto-commits, --no-stream, --no-pretty, --check-update/--no-check-update, --analytics/--no-analytics) were verified against Aider's current CLI options reference rather than assumed.
• No Aider source is copied or vendored — communication is purely via CLI subprocess, same boundary as the Claude Code and mini-swe-agent adapters.