Related: #98748

What Problem This Solves

OpenClaw still has duplicated promise memoizers around lazy dynamic imports in core Gateway and stateful runtime paths. Those local cache owners repeat lifecycle logic, make rejection behavior easier to drift, and increase the cost of maintaining import boundaries.

This is PR 3 of 6 in the lazy-runtime utility consolidation series.

Completed dependency batches:

• #99261 — shared lazy runtime loader foundation.
• #99278 — core leaf lazy loader consolidation.

Remaining batches:

• Batch 4 — Discord, Slack, and Telegram lazy loaders.
• Batch 5 — remaining channel and plugin lazy loaders.
• Batch 6 — residual repository sweep and consolidation cleanup.

Why This Change Was Made

This migrates the core Gateway, session/stateful, task registry, node-host, ACP, channel, and narrow Plugin SDK runtime loaders to the shared lazy-runtime owner. It preserves sticky rejection caching, explicit clear/peek test and lifecycle behavior, startup ordering, and existing dynamic import boundaries without adding SDK exports or compatibility paths.

User Impact

No intended user-visible behavior change. Maintainers get one canonical lazy-loader implementation with less duplicated production code and consistent cache semantics across these runtime paths.

Evidence

• Existing focused proof: 547 tests across ACP, channels, agents, Gateway, SDK, node-host, and tasks.
• Existing focused static proof: core production typecheck, oxlint, and oxfmt passed.
• Fresh rebased build: Blacksmith Testbox-through-Crabbox tbx_01kwjsgnpa39k6r314s21jc6nc, Actions run https://github.com/openclaw/openclaw/actions/runs/28632429729, pnpm build passed.
• Fresh rebased import graph: node scripts/check-import-cycles.ts reported 0 runtime value cycle(s).
• Fresh autoreview before commit: no accepted/actionable findings, confidence 0.94.
• Scope: 23 owned files, 306 insertions and 468 deletions; no changelog entry.