feat: adapter — run evolvent-ai/ClawMark (100 multi-day coworker-agent tasks, rule-based scoring) under the ClawBench harness

Goal

Let the ClawBench harness load and execute tasks from evolvent-ai/ClawMark (100 multi-day, multimodal, multi-environment coworker-agent tasks across 13 professional domains; each task simulates 1–3 working days; strict rule-based Python checkers, zero LLM-as-judge → fully reproducible) so users can:

1. Run a long-horizon, time-staged eval through our 5-layer trace pipeline.
2. Get a corpus with deterministic scoring as a baseline alongside our judge-based two-stage rubric.

Scope

• Stage scheduler — tasks unfold across 1–3 stages (Mon 3/16 → Tue 3/17 → Wed 3/18). Add a stage-advance hook to the harness so the per-day instruction file is loaded and the agent only sees the current day's context.
• Backend mocks — package their backends (filesystem, GreenMail, Notion mock, Google Sheets mock, Radicale CalDAV) as a docker-compose layer the harness can spin up per batch.
• Implicit-state injection — their environment mutates between stages (new email arrives, DB rows update, calendar shifts). Hook this into the existing per-task extra_info/ mount lifecycle so state changes fire at the right boundary.
• Multimodal fixtures — assets/input/ ships screenshots, PDFs, audio, video. Just mount it; existing harness already supports binary fixtures.
• Deterministic scorer — invoke their 10–25 Python checker functions per task and emit a normalized run-meta.json score field. Skip our LLM judge for this corpus (their explicit design choice).
• CLI — clawbench run --corpus clawmark --model <m> end-to-end.

Why now

Their leaderboard already covers the frontier (Sonnet 4.6 at 0.758, Opus 4.6 at 0.746, GPT-5.4 at 0.720). Hosting that corpus under our harness gives our users a deterministic, reproducible alternative metric — and demonstrates that our trace pipeline plays well with rule-based scorers, not just LLM-judge pipelines.

Acceptance

• clawbench run --corpus clawmark --limit 5 --model claude-sonnet-4-6 produces 5-layer trace bundles with stage boundaries marked in actions.jsonl.
• Deterministic checker scores match their upstream scripts/ output exactly on a sampled subset.
• Docs: eval/adapters/clawmark.md + multi-day stage diagram.

Start with the existing per-task extra_info/ mount lifecycle, binary fixture support, and the clawbench run CLI entry point. Compare ClawMark's scripts/ checker output and stage data with the harness trace format, including run-meta.json and actions.jsonl. Done means the stated clawbench command runs five tasks, preserves stage boundaries, matches sampled checker scores, and adds eval/adapters/clawmark.md with the multi-day diagram.