feat(adapters): openclaw adapter

[aktasbatuhan]

Goal

Add an adapter so watchmen captures openclaw sessions alongside Claude Code, Codex, pi.dev, OpenCode, and Cursor (#113). One corpus across tools.

Adapter contract

Per src/watchmen/adapters/init.py, a new adapter is a module exposing:

• NAME: str
• discover() -> Iterable[FileEntry] (walk the install path, yield sessions; return [] when openclaw isn't installed)
• scan(entry) -> (session_dict, prompts, tool_calls)

It registers in the ADAPTERS tuple. claude_code / codex are the cleanest references for a JSONL-on-disk source; cursor.py is the reference for a SQLite key-value store.

Open questions (resolve before coding)

• Where does openclaw persist sessions on disk, and in what format (JSONL per session? a SQLite store? something else)?
• Does it expose per-turn token usage / cost reliably? (If not, leave those at 0 rather than guessing from a price table, like the cursor adapter.)
• Is there a tool-call + outcome signal we can map to is_error / friction signatures (feat: friction ledger — recurring mistakes grouped by error signature (#110) #110)?
• How are skills detected (SKILL.md path heuristic, like the non-Claude adapters)?

Acceptance criteria

• Document the on-disk session format
• Implement discover() + scan() against synthetic fixtures
• Register in ADAPTERS and add to the README agent list
• Unit tests covering the conversation shape(s), tool calls, and skill detection
• Full suite green, no regressions

Note: storage location/format is unconfirmed; first task is to nail that down before writing the parser.