Add Qwen CLI driver

[simone-stacks]

Builds on top of #113 (kimi driver); the diff shrinks to the qwen commits once that merges.

Summary

• Add a qwen-cli driver so swarms can run Qwen Code CLI alongside the existing drivers.
• Install Qwen Code CLI in the agent image from the official standalone archive, with a qwen_cli_version pin.
• Wire the driver through launch, dashboard, docs, and tests.
• Detect token-plan quota exhaustion even when the CLI exits 0, so agents back off and resume instead of dying mid-run.
• Sum kimi token usage from the session wire.jsonl, so the swarmfile pricing map produces real costs for kimi (small fix riding on this branch).

Changes

• lib/drivers/qwen-cli.sh: new driver implementing the swarm interface.
• Dockerfile + launch.sh: install Qwen Code CLI and expose qwen_cli_version; forward QWEN_CONTEXT_WINDOW into agent containers.
• dashboard.sh, USAGE.md, README.md, CHANGELOG.md: surface the new driver.
• lib/drivers/kimi-cli.sh: read token usage from the session wire.jsonl (stream-json carries no usage object).
• tests/: driver/config/dashboard/harness/launch coverage and qwen fixture swarmfiles.

Self-review checklist

• This PR addresses a single concern (why?). If it covers multiple independent changes, I've split them into separate PRs.
• The new driver is covered by unit tests.
• The quota detection and wire-usage paths are covered by unit tests.
• shellcheck -s bash passes on all modified .sh files.
• New lines match the surrounding style (these files already have pre-existing >79-col lines).

Test plan

• shellcheck -s bash lib/drivers/qwen-cli.sh lib/drivers/kimi-cli.sh dashboard.sh launch.sh tests/test_config.sh tests/test_dashboard.sh tests/test_drivers.sh tests/test_harness.sh tests/test_launch.sh
• tests/test_drivers.sh (557 passed)
• tests/test_config.sh (375 passed)
• ./tests/test.sh --unit (note: test_harvest.sh has pre-existing failures on this machine — signing env, green in CI)
• Live runs on stacks-swarm (9-agent validation + 4-agent engagement on qwen3.8-max): real token counts in the dashboard, quota backoff triggered instead of agent death, one triaged finding produced.