fix(coding-agent): stop retry watchdog after provider recovery

[asdfqwerzxcc]

Summary\n\nStop the provider-timeout retry watchdog from aborting a recovered run after the provider has already produced a tool call. The watchdog now checks retry-request ownership before aborting the active Agent, so a long foreground eval can cross the old deadline and still return its result to the model.\n\nThis is the separate lifecycle defect observed while reproducing oh-my-openagent #8246: no Esc/Ctrl+C input was involved.\n\n## Changes\n\n- add a retry-request ownership predicate to the bounded continuation watchdog\n- derive ownership from AgentSession retry-attempt state, which resets at the recovered provider response boundary\n- add deterministic fake-timer coverage for provider timeout -> recovered eval -> old deadline -> provider call three\n- preserve the existing watchdog for genuinely wedged provider retries\n- document the core ownership contract and release note\n\n## QA & Evidence\n\n- RED: parent Agent signal became aborted and only 2 provider calls occurred\n- GREEN focused: 11/11 timeout continuation tests pass\n- Provider regression cluster: 32/32 pass\n- Static/build: changed-file Biome pass, root TypeScript no-emit pass, diff check pass, full build exit 0, changelog gate pass\n- Manual runtime driver: recoveredParentAborted=false; a still-wedged retry remains wedgedRetryAborted=true\n- Real source CLI: eval hard-limit mock loop 5/5 pass, two model turns served, final assistant text returned, real auth unchanged\n- Artifact: local-ignore/qa-evidence/20260914-retry-watchdog-tool-abort/README.md in the task worktree (ignored by repository policy)\n\n## Risks & Residuals\n\n- The predicate relies on the existing invariant that a genuinely successful provider message_end resets retry-attempt state before tool execution. The regression pins that ordering.\n- Broad Windows suites retain pre-existing path, release-script, and native-prebuild fixture failures documented on the sibling PR; focused changed-domain tests are green and GitHub CI is authoritative.\n- The local gate-review agent remained non-terminal in its bounded attempt, so no local reviewer approval is claimed; repository CI/review remains required.\n\n## Related\n\n- Sibling file-integrity fix: #1670\n- Original report: https://github.com/code-yeongyu/oh-my-openagent/issues/8246\n

---

Summary by cubic

Stops the provider retry watchdog from aborting a session after a timeout retry has already recovered. Previously, a retry that successfully yielded a long-running local tool could still hit the original watchdog deadline, abort the Agent mid-execution, and drop the tool result. Now the watchdog only aborts while a retry request is still pending, so a recovered run continues normally.

• The pending check is derived from retry-attempt state, which resets when the recovered provider response reaches message_end before the tool runs.
• Adds a deterministic fake-timer regression test covering the timeout → recovered eval → old deadline → provider call 3 sequence.
• A genuinely wedged provider retry is still aborted by the watchdog as before.

^{Written for commit 168f2ee. Summary will update on new commits.}