Add Claude Code CLI as a ringer engine

[hcoronel1]

Summary

• Adds the claude/claude-sonnet engine identity and model registry entries (claude-sonnet-5, claude-opus-5, claude-fable-5-1, claude-haiku-4-5, alias sonnet) so Claude Code CLI can run as a worker engine alongside Codex, Grok Build, and OpenCode/OpenRouter.
• Adds result_parser = "claude-code-json" and parse_claude_code_json_result() to parse the claude -p --output-format json result envelope: real token usage from usage, and the effective model from modelUsage (Claude Code can resolve an alias like sonnet to a dated model id).
• Distinguishes transport/API failures (rate limits, timeouts — marked scoreable=False, failure_kind="infrastructure", excluded from the scoreboard, no verify/retry) from agent-level failures (max turns, tool/permission rejection — still scoreable=True, so a genuine quality loss isn't hidden from the scoreboard). An earlier version of this PR classified every is_error as infrastructure; fixed after a Codex read-only audit caught it.
• Reads the JSON result structurally — one JSON object per line, or the whole trimmed payload as a single object — instead of scanning the text for a bare {. The earlier bracket-scan could match a { an agent happened to quote inside its own result text and misattribute tokens/model/error to that embedded fragment. Also fixed after the same audit.
• _claude_reported_model only returns a family match (sonnet/opus/fable/haiku) when every same-family candidate in modelUsage agrees on the same canonical model; otherwise it returns None instead of guessing the first one. Guards against misattributing identity when a director and a same-family subagent both appear (e.g. two different Sonnet versions).
• Schema v3→v4 migration (scoreable, failure_kind columns) with a backward-compatible ALTER TABLE.

Motivation

I wanted to run real production tasks through Ringer using Claude Code CLI as the worker, and needed the scoreboard to stay honest about why a run failed — an API rate limit isn't the model's fault, but Ringer's existing token/model regex parsers had no way to tell an infra failure apart from a real one for a JSON-emitting engine.

Test plan

• python3 -m pytest -q — 278/278 passed (was 275 before this branch's own fixes)
• New coverage in tests/test_claude_engine.py: JSON result selection (including the embedded-JSON-in-string case), infra vs. agent error classification, ambiguous same-family model attribution, schema migration
• Self-reviewed with a Codex read-only audit before opening this PR; findings incorporated (see bullets above)

🤖 Generated with Claude Code