Fix backend runner lifecycle handling

[DorianZheng]

Summary

• Delay JSONL CLI host result emission until backend completion instead of treating intermediate assistant messages as final
• Track live BoxLite boxes with atomic heartbeat records so startup cleanup preserves active runs
• Force-remove boxes during explicit agent deletion and clear heartbeat records on teardown and error paths
• Add regression coverage for message completion handling and box ownership cleanup races

Why

The runner could finalize too early on the first assistant message, and startup cleanup could remove an active VM when AgentLite instances overlapped. The new heartbeat flow distinguishes live owners from stale boxes, including PID reuse, failed removals, and partial-write cases.

Impact

Active runs are less likely to be interrupted, explicit deletes still clean up boxes, and host clients receive a single final result after backend completion.

Validation

• PATH=/opt/homebrew/bin:$PATH ./node_modules/.bin/tsc --noEmit
• git diff --check
• PATH=/opt/homebrew/bin:$PATH ./node_modules/.bin/vitest run --maxWorkers=4 (one timing assertion in src/acp/client.e2e.test.ts failed during the full parallel run)
• PATH=/opt/homebrew/bin:$PATH ./node_modules/.bin/vitest run src/acp/client.e2e.test.ts --maxWorkers=1
• PATH=/opt/homebrew/bin:$PATH ./node_modules/.bin/vitest run src/box-runtime.test.ts src/container-agent-backend.test.ts src/container-runner.test.ts --maxWorkers=4