fix(core): storage hardening — Buffer safety, embedding abort, atomic instanceId

[rainforest888]

fix(offload): Tracing & Token Hardening

拆分自 #391，per YOMXXX review。

修复范围

聚焦 src/offload/ 层：token estimate、local LLM、Opik tracing。不涉及 store、runtime、gateway 层。

具体修复

src/offload/fast-token-estimate.ts — Astral-平面 Token 修正

问题	修复
astral-plane 字符（emoji / CJK extension B，U+10000+）编码为 surrogate pair，原实现把 high surrogate 和 low surrogate 各计为一个 token	high surrogate 检测后跳过低代理（i += 2），修正 ~2× token 高估

src/offload/local-llm/index.ts — L1.5 Fallback 修正

问题	修复
解析失败返回 {taskCompleted: false, ...} — normalizeJudgment 把 false 当有效判断，永远不触发 "LLM 不可用" 重试	改为 {taskCompleted: null, ...}，== null 检查触发重试路径

src/offload/opik-tracer.ts — ESM 兼容 + Trace 截断

问题	修复
ESM 下 require("opik") 为 undefined，tracer 永久静默关闭	改为 createRequire(import.meta.url)
未脱敏 tool 输出（文件内容、环境变量）完整写入 trace，泄露到 Opik backend	统一截断到 2000 字符

验证

• 3 files, +41/-15
• 无 API 变更，无新增依赖
• 与 fix(gateway): harden CORS / Host header / startup host binding #39 / fix(tcvdb): escape session filter literals #232 / fix(offload): keep backend TLS verification enabled #242 / fix(offload): sanitize session storage paths #287 / fix(offload): keep MMD files inside storage dir #288 / fix(offload): sanitize ref and skill paths #289 / fix(offload): invalidate token cache after tool use stripping #347 均不重叠

[Maxwell-Code07]

Thank you for your contribution! We will review your code and get back to you as soon as possible.

[YOMXXX]

Thanks for the focused core hardening split. I verified the branch locally with:

• pnpm test — passed: 4 files, 67 tests
• pnpm build — passed
• git diff --check HEAD~2..HEAD — passed
• a local Float32Array subarray Buffer check — passed

I think this still needs one change before approval: the new EmbeddingCallOptions.abortSignal is wired into the embedding HTTP fetch, but the actual auto-recall timeout path does not pass an abort signal. performAutoRecall() still uses Promise.race with a timer that resolves after recall.timeoutMs; once that race resolves, performAutoRecallInner() and the embedding HTTP request can continue running until the embedding request's own timeout. With defaults, recall.timeoutMs is 5000ms while embedding.timeoutMs is 10000ms and embedding.recallTimeoutMs may be unset, so the PR does not fully deliver the documented "recall timeout aborts the in-flight embedding request" behavior.

Please create an AbortController for the recall timeout path and pass its signal through searchMemories() / EmbeddingCallOptions so the same timeout that skips memory injection also cancels the embedding fetch. The Buffer byteOffset/byteLength fixes and the instance-id in-flight guard look directionally good.