test(recall): cache accounting benchmark for split_system (SIMULATION)

[muzimu217]

摘要

为 #449（split_system 把稳定 persona 前置到 CACHE_BOUNDARY）补充一个缓存记账基准测试，验证 split_system 确实增大了"缓存边界之前"的可缓存前缀长度。

⚠️ 本文档/测试是确定性仿真，不是真实 A/B。
它不调用任何 LLM API，不测量真实 cache_read / cache_creation token。
真实测量方法请见社区 PR #433（2 provider × 2 variant × 2 repeat × 3 turn = 24 个实测 cell）。本 PR 只做两件事：① 文档化 split_system 的预期行为；② 防止前缀组装回归。

为什么需要这个测试

前缀匹配缓存（DeepSeek / MiMo / Anthropic 类）只对 CACHE_BOUNDARY 之前、且逐 token 完全一致的前缀做缓存。split_system 的价值 = 把唯一稳定的 persona 移进这段可缓存区域。本测试断言的唯一不变量：

• split_system 开启时，缓存前缀长度 严格大于 none；
• persona（稳定内容）在边界前；动态 L1 记忆（每轮变化）不在边界前。

测试内容

src/adapters/openclaw/cache-accounting.benchmark.test.ts（4 用例）：

1. split_system 缓存前缀含 persona 且长于 none
2. 每轮 cache_read 增益可观且确定性（约 10–40 token）
3. 20 轮会话：split_system 累计 cache_read 严格更大，且增益随轮数线性（无随机）
4. persona 是唯一移入边界前的内容；动态 L1 标记（<relevant-memories> / <memory-context>）不应出现在边界前

与 #449 / #375 / #433 的关系

• fix(recall): move persona+scene to prependSystemContext — core prefix stability for #120 #449：本测试验证的实现对象（split_system 前缀组装）。
• fix(recall): stabilize prompt-cache memory injection #375：canonical runtime fix（解决动态 L1 位置），与本测试正交。
• fix(recall): move dynamic L1 memory to appendContext to preserve prefix cache #433：真实 provider 实测方法论来源——本测试明确引用其作为"真实数字"的权威，不自称已测得真实命中率。

Scope / Non-scope

• 本 PR 做：一个诚实标注的仿真基准 + 回归守卫。
• 本 PR 不做：真实 A/B 测量（应由 fix(recall): move dynamic L1 memory to appendContext to preserve prefix cache #433 式方法在真实 provider 上完成）；不修改任何生产代码。
• 本 PR 不声称：split_system 有"X% 命中率提升"——那是 fix(recall): move dynamic L1 memory to appendContext to preserve prefix cache #433 的实测范畴，本测试只验证前缀组装逻辑正确。

[YOMXXX]

Duplicate/scope cleanup triage:

This is useful coverage, but it should not stay as a standalone PR. The test file only validates the split_system behavior introduced by #449, and the PR body explicitly treats #449 as the implementation under test. If #449 is accepted, this benchmark/regression test should be folded into #449 so the behavior and its guard land together. If #449 is not accepted, this PR has no independent production behavior to protect.

I would close this as a separate PR and move the test into #449 during the next revision. The real-provider measurement lane should remain separate from this simulation, as noted in the PR body.

[muzimu217]

This test suite has been superseded by improved tests in #449.

Issues with this PR:

• Tests were tautological (testing own mock)
• Did not verify real host behavior
• No value as standalone PR

Replacement:

• Better tests added to fix(recall): move persona+scene to prependSystemContext — core prefix stability for #120 #449 (src/adapters/openclaw/auto-recall-cache.test.ts)
• Real host contract verification tests
• Honest documentation of unverified assumptions

Closing to consolidate testing in the main PR.