feat(openclaw): assemble auto-recall context server-side

[starslittle]

Summary

• replace OpenClaw auto-recall's client-side find/rerank/read path with one session-aware server context search
• preserve target selection, actor/tenant routing, injection markers, timeout controls, and recall traces while enabling query expansion and five-turn deduplication
• map existing recall limits and character budgets onto the server context contract, and update the related tests and documentation

Testing

• npm run typecheck
• npm run build
• focused Vitest contract suite: 143 tests passed
• remaining plugin regression suite: 42 files, 667 tests passed
• npm pack --dry-run --json

Refs #4415 (item 2)

[ZaynJarvis]

One concrete follow-up after reviewing this PR: please reuse or extend the shared context-search contract instead of maintaining a separate OpenClaw searchContext request body.

References:

• examples/memory-plugin-shared/lib/recall-core.mjs — buildContextSearchBody, contextRequestTimeoutMs, fetchAssembledContext, and normalizeContextEntry
• examples/claude-code-memory-plugin/scripts/auto-recall.mjs — current consumer
• examples/pi-coding-agent-extension/shared/recall-core.mjs and recall-core.d.mts — vendored MJS + TypeScript typing pattern

The OpenClaw-specific context_type / detail, routing, and trace handling can remain a thin adapter around that shared contract. A request-parity test would help prevent the plugin implementations from drifting again.

[starslittle]

Thanks for the concrete follow-up. I updated the branch to vendor the canonical memory-plugin-shared recall core into OpenClaw via the existing sync mechanism. OpenVikingClient.searchContext now derives its base request from buildContextSearchBody, uses contextRequestTimeoutMs, and normalizes returned entries with normalizeContextEntry. OpenClaw-specific context_type, detail, tenant/actor routing, logging, and recall traces remain thin adapters.

I also added a request-parity test against the shared builder, session-aware timeout and response-normalization coverage, an end-to-end request assertion, and package/install checks ensuring the shared runtime ships in both source and dist.

Validation completed: npm run typecheck, focused Vitest (98 tests), remaining plugin regression suite (42 files / 670 tests), shared recall/sync tests, npm run build, compiled-runtime import smoke test, and npm pack --dry-run --json. The full Windows run retains the repository’s existing six architecture-boundary failures and one WSL path failure; no new failures were introduced.