jiuwenclaw adapter: adapt to upstream rename + fix subprocess timeout hang

[lec77]

Linked issue

Closes #8.

Summary

Adapt the jiuwenclaw adapter to upstream's package rename (jiuwenclaw → jiuwenswarm, v0.2.0) and the wire-format updates that followed, and root-cause-fix an 8.75h subprocess hang we hit during testing where a client-side timeout couldn't kill the spawned ACP child cleanly.

Changes

• Adapt to upstream wire/package changes — rewrite the history.json parser for the current flat per-event shape; switch console script (jiuwenclaw-cli → jiuwenswarm-tui), module paths (jiuwenclaw.app → jiuwenswarm.app, jiuwenclaw.app_cli → jiuwenswarm.channels.acp.app_acp), runtime dir (~/.jiuwenclaw/ → ~/.jiuwenswarm/), and the pkill pattern. SkVM adapter name stays jiuwenclaw so CLI flags (--adapter=jiuwenclaw), config key, and ~/.skvm/proposals/jit-optimize/jiuwenclaw/... paths remain stable.
• Fix subprocess timeout hang (shared runCommandWithEnv in hermes.ts) — spawn with detached:true and group-kill on timeout, plus a bounded pipe-drain so a surviving grandchild can't deadlock the EOF read. The prior code SIGTERM'd only the direct child; the jiuwenswarm ACP client spawns a grandchild that inherits the stderr pipe, so the EOF read never completed.
• Stop server-side runaway on timeout (jiuwenclaw adapter) — on timedOut, restart the long-lived sidecar so a runaway session in the AgentServer can't keep burning CPU / poison the next run sharing the singleton sidecar.
• Pass realpath of task.workDir — macOS mkdtemp returns /var/folders/... symlinks; resolving once before threading into the prompt hint and --workspace-dir keeps the hint, inputs["cwd"], and the agent's emitted absolute paths consistent.

Test plan

• bunx tsc --noEmit
• bun test — 103 adapter tests pass (regression suite for hermes covered too)
• Manual verification:

# A. Timeout no-hang
bun run skvm run --task=skvm-data/tasks/weather_task_01/task.json \
  --adapter=jiuwenclaw --adapter-config=managed \
  --model=deepseek/deepseek-chat --timeout-ms=20000
# → returns in ~33s (20s timeout + ~6s sidecar restart), runStatus=timeout

# B. Bench end-to-end across pdf+weather × [no-skill,original]
bun run skvm bench \
  --tasks=pdf_task_01,pdf_task_02,weather_task_01,weather_task_02 \
  --model=deepseek/deepseek-chat \
  --adapter=jiuwenclaw --adapter-config=managed \
  --conditions=no-skill,original --concurrency=1
# → exit 0, all 8 runs complete

# C. jit-optimize end-to-end with optimizer engagement
bun run skvm jit-optimize \
  --skill=skvm-data/skills/weather --task-source=real \
  --tasks=weather_task_01,weather_task_02 \
  --target-model=deepseek/deepseek-chat --target-adapter=jiuwenclaw \
  --optimizer-model=cheap_ipads/gpt-5.5 \
  --adapter-config=managed --rounds=1 --runs-per-task=1 --max-steps=80
# → exit 0, optimizer engages and writes proposal

Notes for reviewers

• The PR scope grew beyond issue jiuwenclaw adapter: support per-request workspace_dir + chat.usage from upstream #8 during implementation: while running skvm bench to verify the wire-format work, we hit an 8.75h subprocess hang (commits 3 + 4). Both are bug-fix-shaped under the CONTRIBUTING.md rule, so no separate issue was filed — happy to file one retroactively if preferred.
• Two of the four commits touch shared code (runCommandWithEnv in hermes.ts) — strictly safer (kills more, doesn't hang on EOF), and the hermes regression suite passes, but worth a second look from someone familiar with hermes.
• Companion upstream PR at openJiuwen-ai/jiuwenswarm#34 adds the protocol-side features this adapter consumes. That PR is independently mergeable; this adapter ships against a jiuwenswarm that has those features.
• Known limitation (out of scope): an unrelated jiuwenswarm-side issue where write_file silently no-ops on some file-output tasks. To be filed upstream separately.

Checklist

• Linked issue jiuwenclaw adapter: support per-request workspace_dir + chat.usage from upstream #8
• Commit messages follow <scope>: <summary> style
• Added/updated tests where appropriate
• Updated docs/jiuwenclaw.md to reflect upstream rename + new path
• No unrelated changes mixed in