obs: per-request [acp-usage] on the plugin path + fold-materialization anchor ceiling

[ranxianglei]

Fixes #695

背景

用户报告 plugin 会话「压缩后缓存掉 0」。现场调查(见 issue)只能给出理论上限 ≈15–18%(系统提示+块摘要),无法区分折叠物理与上游逐出,因为 plugin 路径没有任何 per-request 日志(wire 路径有 [acp-usage],plugin 路径静默更新内存)。

改动(纯观测,零行为变化)

1. plugin 路径 per-request [acp-usage](src/plugin.ts applyUsageSample,现导出供测试):
   [<sid>] [plugin] [acp-usage] input=50000 cached=8000 (cache hit 16%) ctx=51200 fold=new
   • fold=new 精确标出折叠落地的那一个请求(applyRanges 置位 → 首个 usage 报告消费)
   • cached 缺失时打 cached=n/a,不算命中率
2. [acp-compress-obs] 扩展锚点估计(src/stream.ts):
   ... anchor≈1234 tok (3 active blocks, sys excluded) postCtx≈31000 → next-request cache ceiling ≥3%
   • anchor = Σ active 块 summary chars/4(下界,sys 未知)
   • 下一请求真实 cached ≈ ceiling → 物理保留;≪ ceiling → 上游逐出/争用
3. wire 路径同标 fold=new(src/loop/core.ts)保持两条路径对称。
4. Session.stats.pendingFoldUsage?: boolean(src/session.ts),内存态,不落盘。

验证

• 新增 tests/usage-observability.test.ts 4 项:置位/锚点行字段、usage 行格式+fold 消费、无残留标记、cached=n/a
• 全套 1322 pass / 0 fail(2 skip),tsc + build 绿

[github-actions]

📦 Built Package Artifact

Branch: obs/plugin-cache-usage-695 (d6b955a)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-696

Each push to this PR publishes a new version under the pr-696 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr696.tgz
npm install -g package

---

This comment is automatically updated on each push.