fix: handle stale web observe runners (#875)

Co-authored-by: Codex <codex@noreply.local>

scripts/src/hwlab-node-help.ts

@@ -84,6 +84,7 @@ export function hwlabNodeWebProbeHelp(): Record<string, unknown> {
       "bun scripts/cli.ts hwlab nodes web-probe observe command webobs-xxxx --type steer --text '继续观察当前 trace'",
       "bun scripts/cli.ts hwlab nodes web-probe observe command webobs-xxxx --type cancel",
       "bun scripts/cli.ts hwlab nodes web-probe observe status webobs-xxxx",
+      "bun scripts/cli.ts hwlab nodes web-probe observe stop webobs-xxxx --force",
       "bun scripts/cli.ts hwlab nodes web-probe observe collect webobs-xxxx --view turn-summary",
       "bun scripts/cli.ts hwlab nodes web-probe observe collect webobs-xxxx --view trace-frame --trace-id trc_xxx --sample-seq 42",
       "bun scripts/cli.ts hwlab nodes web-probe observe analyze webobs-xxxx",
@@ -100,9 +101,11 @@ export function hwlabNodeWebProbeHelp(): Record<string, unknown> {
       "Issue-ready evidence is available under issueEvidence and summary.issueEvidence; full script report is persisted under probe.reportPath with a SHA-256 fingerprint.",
       "observe sampling is passive by default: it records DOM summaries and natural page request/response/requestfailed events with observerInitiated=false; it does not actively fetch Workbench APIs, reload, switch sessions, route/intercept, or call repair helpers.",
       "observe start registers a local UniDesk-side observer id under .state/web-observe/index.json; after start, prefer observe status|command|stop|collect|analyze <id> instead of repeating --node/--lane/--state-dir.",
+      "observe status reports runner liveness separately from process existence, including heartbeat age, stale threshold, command backlog, and abandoned command counts.",
       "observe command actions are explicit user/control actions and are appended to control.jsonl; use --type newSession/selectProvider/sendPrompt/steer/cancel/goto/screenshot/mark/stop. steer/cancel reuse the Workbench composer path, not private backend APIs. For long prompts use sendPrompt/steer --text-stdin; keep prompt text out of issue comments by citing textHash/textBytes.",
+      "observe stop --force first checks heartbeat/backlog health; if the runner is stale or commands are not being consumed, it kills the recorded PID from outside the command queue and marks pending/processing commands abandoned so analyze classifies them as tooling findings.",
       "observe collect --view turn-summary renders the multi-turn CLI reading layer from samples/control artifacts; --view trace-frame renders one sampled trace frame with a fixed Final Response block. collect views do not save a second source of truth.",
-      "observe analyze is offline-only: it reads artifact JSONL and writes analysis/report.md plus analysis/report.json without accessing Workbench APIs or driving the browser.",
+      "observe analyze is offline-only: it reads artifact JSONL plus observer command/heartbeat artifacts and writes analysis/report.md plus analysis/report.json without accessing Workbench APIs or driving the browser.",
       "observe analyze scans every sampled DOM point, extracts Workbench timing text such as 总耗时/total and 最近 N 秒/分前, and writes a sample point vs turn timing report: each Markdown table row starts with the timestamp, followed by each turn's 总耗时(s) and 最近更新(s). Timing series are reported for post-processing/manual analysis instead of auto-judged from status tail output.",
       "observe analyze also reports visible “加载中” count, owner attribution, concurrent loading owners, and continuous visible segments; fixes must reduce real loading latency, not reveal incomplete content early to make this metric disappear.",
       "script/observe support --browser-proxy-mode auto|direct; use direct for A/B evidence when frontend RUM is slow but OTel server spans are absent or fast, instead of falling back to raw Playwright.",