fix(workflows): honor Pi effective HTTP idle timeout for child progress watchdog

[seekskyworld]

Problem

Workflow child sessions currently derive MODEL_PROGRESS_TIMEOUT_MS from the raw optional httpIdleTimeoutMs fields in project/global settings. When the user has no explicit override, those fields are absent even though Pi's effective setting is its 300-second default. OpenPI therefore applies an undocumented 45-second provider-progress watchdog to every child, which can abort slow but healthy provider turns before Pi's own transport timeout.

This is a focused reliability slice of #424. It does not change tool execution, provider protocol support, or cancellation semantics.

Evidence

• Pi 0.85.1 exposes SettingsManager.getHttpIdleTimeoutMs(), returning the effective default of 300000 ms.
• extensions/workflows/runner.ts currently reads getProjectSettings()/getGlobalSettings() directly and falls back to 45000 ms when both raw fields are absent.
• The existing watchdog remains useful as a bounded child guard, so an explicit Pi timeout of 0 will retain the OpenPI minimum rather than creating an unbounded child.

Acceptance

• Resolve the watchdog default from Pi's effective HTTP idle timeout API.
• Preserve the 45-second minimum for effective values below that floor, including explicit 0 (disabled transport timeout).
• Preserve test-only overrides and existing abort/cleanup behavior.
• Add deterministic regression coverage for the effective default, wider values, floor values, and explicit zero.
• Update the user-facing timeout contract and link this issue to fix: make child tool, workspace and provider execution reliable #424.

[seekskyworld]

Implemented in PR #427. The Workflow child progress watchdog now uses Pi’s effective SettingsManager.getHttpIdleTimeoutMs() value, keeps the 45-second floor (including explicit 0), and preserves test overrides and abort/cleanup behavior. Validation and evidence boundary are recorded in docs/research/ISSUE_425_WORKFLOW_WATCHDOG_2026-09-07.md. This closes only the focused reliability slice claimed from #424.

[seekskyworld]

上游已合并 #426（https://github.com/openpi-dev/openpi/pull/426，commit eaf470b）。该更宽的修复移除了额外的子会话输出 watchdog，统一由 Pi 原生传输超时与重试负责，因此已覆盖本 Issue 的 45 秒默认超时问题。关联 PR #427 已按重复实现关闭；本 Issue 不再需要独立代码变更，现关闭并保留研究记录与验证提交作为追溯证据。