ci(windows): require crash recovery evidence

[liugddx]

Summary

Add a required windows_recovery job that turns the Windows crash evidence landed across #2475, #2488, #2492, and #2509 into one bounded release gate.

The gate builds on Node 24 and separately verifies SQLite transaction recovery, Runtime resume/continuation recovery, Runtime Host owner-death recovery, and managed-workspace crash convergence. A workflow contract test prevents the job from becoming non-blocking or silently dropping an evidence surface.

Raise only the Windows timeout for the two Runtime crash harnesses from 60 to 120 seconds. The Phase 0 harness serially launches 12 real child processes and exceeded the old aggregate timeout locally; POSIX retains the existing 60-second boundary.

Refs #2142

Verification

ode --test scripts/windows-recovery-workflow.test.mjs (1 pass)

• Biome check for the changed TypeScript/JavaScript files
• Parsed .github/workflows/windows-recovery.yml with the repository YAML dependency
• git diff --check
• SQLite crash recovery command: 11 pass, 0 fail
• Runtime continuation Phase 1 passed locally; Phase 0 reached its old explicit 60-second timeout, which this PR raises on Windows
• Full local rebuild was blocked by the existing installed ypescript@7.0.2 package not exposing the compiler API expected by @maka/code-mode; the clean Node 24 workflow run is authoritative
• Existing locally built Runtime Host artifacts could not start their fixture Host and the managed-workspace command exceeded the five-minute tool limit; both are rerun from a clean build by the new job

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes - Windows recovery evidence becomes release-blocking.
• No

[liugddx]

Adversarial review follow-up / 对抗性审查后续

EN:

• Hardened the two name-filtered recovery commands with TAP evidence: Runtime Host must report exactly 2/2 passing with 0 skipped; managed workspace must report exactly 12/12 passing with 0 skipped. A renamed test can no longer make the gate pass with zero matches.
• Removed the source-code timeout regex test; the real Windows harness is the evidence.
• Removed duplicate SQLite and managed-workspace crash runs from the non-blocking baseline. The required recovery lane now owns crash evidence; baseline keeps path/lock, PTY, smoke, and nightly full-storage evidence.
• Windows recovery passed end to end in 3m39s on commit 80c9317.
• Maintainer check before merge: confirm windows_recovery is required by the repository ruleset (the current token cannot read that setting).

中文：

• 为两个按名称筛选的恢复命令增加 TAP 精确计数：Runtime Host 必须 2/2 通过且 0 跳过，managed workspace 必须 12/12 通过且 0 跳过，测试改名后不会出现“零匹配假绿”。
• 删除对源码超时常量的正则快照；真实 Windows crash harness 才是有效证据。
• 从非阻塞 baseline 删除重复的 SQLite 与 managed-workspace crash 运行，由 blocking recovery lane 统一负责；baseline 继续负责路径/锁、PTY、启动 smoke 与 nightly full-storage。
• 提交 80c9317 的 Windows recovery 已完整通过，耗时 3 分 39 秒。
• 合并前请 maintainer 确认仓库 ruleset 已将 windows_recovery 设置为 required check；当前 token 无权读取该配置。