fix: let slow auto-compaction run to its configured timeout instead of stopping at 6 minutes (#201928)

Root cause: `src/sessions/session-controller.watchdog.ts:103` (base `0eb6488cd3`) builds the reply watchdog's semantic deadline from the last progress plus 6 minutes and the open model requests only, so a nested compaction attempt's own deadline (`compaction.timeoutSeconds`, set in `compaction-session-agent.ts:221`) never counts and the turn is stopped as stalled once a long compaction summary has run past 6 minutes.

Fixes #201576

Thanks @kazuyuki-eguchi for the detailed report and timeline, and @chessing234 for the first fix in this PR (its two checks are kept as one regression in the watchdog suite).

## What Problem This Solves

With `agents.defaults.compaction.timeoutSeconds: 600`, auto-compaction on a slow local model was cut off after about 6 minutes with `Reply operation stalled`, so the 600 s budget could never be used. The turn failed, the next turn hit the same overflow, and the cycle repeated.

## User Impact

User impact: a quiet auto-compaction can now run until its own `compaction.timeoutSeconds`, and the turn continues once the summary arrives. Stop still cancels compaction at once, compaction still ends at its own timeout, and an ordinary stalled model call is still stopped at the usual window.

## Why This Change Was Made

Compaction already registers with the reply watchdog as a nested maintenance attempt with its own deadline (refreshed while it makes progress). The watchdog honoured that deadline only as a hard stop, not as a reason the turn was not stalled. It covered the summary request only while it was open, measured from the last progress before compaction began. Now a live nested attempt's deadline counts toward the stall window, and when compaction finishes the stall window restarts from that point. Native runtime-owned waits already work this way. No new setting, timeout or wording.

The earlier version of this PR added a separate `compaction-watchdog-allowance` module around recovery compaction. It is removed: the watchdog owner already has the hook, and this covers every compaction that runs through the compaction session, not only recovery.

## Evidence

Isolated Gateway (`ocx src`, pinned Bun, task-owned state), reporter's compaction config (`safeguard`, `qualityGuard`, `midTurnPrecheck`, `timeoutSeconds: 600`), and a scripted local OpenAI-compatible model. A turn reads large files, mid-turn context pressure triggers auto-compaction, and the compaction summary request stays quiet. Only the Gateway process's clock is stepped. Observed lines: https://github.com/openclaw/openclaw/pull/201928#issuecomment-7017676022

| Case | Base `0eb6488cd3` (main) | This PR |
| --- | --- | --- |
| Quiet summary, answered after 420 s | `lane task error: lane=main durationMs=427663 error="Reply operation stalled"`, `caller_signal_aborted`; user sees "Your request couldn't be completed." | `auto-compaction succeeded ...; retrying prompt`, turn answers |
| Quiet past 600 s | n/a | compaction ends at its own timeout (`[compaction-diag] fallback ... reason=timeout`), turn continues |
| User Stop during compaction | n/a | `chat.abort` -> compaction ends 190 ms later, `agent run aborted` |
| Ordinary model call quiet 370 s, no compaction | request closed at the window, then retried | same |

Regression: `session-controller.watchdog.test.ts` "lets quiet compaction outlast the stall window while Stop still cancels it" fails on base (stopped as stalled at 7 min) and passes here.

Change size against the merge base: production +16/-2 (one file), tests +19. The contributor's new module and its test (+89) and the `compaction-runtime.ts` wrapper are gone.

Merge danger: two-way door, no stored state or config change; revert restores the old window. Blast radius: replies whose turn runs compaction (embedded runtime), and only while compaction runs; ordinary model and tool stalls are unchanged.

Co-authored-by: Ayaan Zaidi <hi@obviy.us>