feat(openclaude): OpenClaude CLI provider (#213)

[ozymandiashh]

Implements #213, with @iamtoruk's May yes standing as the feature sign-off (offered in the issue thread tonight; kick it back if that reading was too generous).

What OpenClaude turned out to be (this is why the provider is small): the active distribution is @gitlawb/openclaude on npm, and it is a Claude Code fork. Transcripts are Claude-Code-schema JSONL at ~/.openclaude/projects/<slug>/<uuid>.jsonl, tool names are already codeburn-canonical, usage is Anthropic-shaped. The provider therefore follows the cline-cli reference shape with no tool mapping table.

Semantics: only usage-bearing assistant lines become calls; queue-operation/last-prompt bookkeeping and replay.json siblings are skipped; isSidechain: true subagent lines are counted (real spend); no cost field exists in the transcript, so every call is priced through the shared tables with costIsEstimated: true; project = basename of the first cwd, slug fallback; CODEBURN_OPENCLAUDE_DIR overrides the root; probeRoots() reports the projects dir per the NEW_PROVIDER checklist.

Real local testing proof (CONTRIBUTING bar): I installed the actual CLI and generated real sessions against DeepSeek (--provider openai + base URL https://api.deepseek.com/v1, model deepseek-chat), then ran the provider end to end on them:

sources: 4, calls: 13, cost: $0.0062
models: [deepseek-chat, <synthetic>]
projects: [demo-proj, second-proj]   <- basename of real cwds
probeRoots: ~/.openclaude/projects (projects)

Honest limitation, documented in docs/providers/openclaude.md: streamed responses can split tool_use blocks across assistant events and only usage-bearing events are parsed, so tool breakdowns are a floor. Token and cost accounting are exact.

18 new tests (discovery incl. replay-skip, parsing, dedup across files sharing a session_id, sidechain counting, project resolution, cost estimation), registry updated, tsc --noEmit clean, 36/36 across the two touched suites.