# Validation — 验证事实 / Verified facts

De-identified facts only: no real session ids, private paths, fixture keys, or
checkpoint hashes appear here. These results are NOT a proof of lossless recall
or of superiority over any other compaction implementation.

本文件只含去标识化事实：不含真实会话 ID、私密路径、夹具密钥或 checkpoint 哈希。
这些结果**不是**无损回忆的证明，也不构成相对任何其他压缩实现的优劣证明。

## 0.3.3 image replay repair / 图片回放修复

The complete image repair was transplanted onto published compaction `v0.3.2`
(commit `48762d23b83c19fe517d319404ce0007c933e7e9`) in an independent release checkout.
The ordinary runtime source is unchanged from the accepted local R2; release version,
installer default source and documentation are separate packaging changes.

- `npm run check`: **163 plugin + 46 legacy-A + 34 comparison = 243 passed**.
- Explicit isolated paired suite against the exact published Accounts `v5.1.3`
  commit `dddb7f2bd62ddffc24c62916b205796b1c78c5a0`: **275 account + 27 integration
  passed**, 1 existing optional 130-second real-time case skipped. Virtual-clock
  setup/idle/total deadline tests passed. Total: **545 passed, 1 skipped**.
- The earlier image-development pairing included unrelated Antigravity fixes and
  had 279 account tests; it is not substituted for the exact-tag pairing above.
- Four real temporary `LocalAttachmentStore` cases exercise user/tool images in
  ordinary generation and Basic compaction. Strict fake-service and aggregate-bound
  tests retain stock Pi defaults and rejection guards. Real-owner integration checks
  wire images plus native state, Basic readable-summary commit/restore, and no history
  replacement or stock retry on reader failure. Transport/auth are injected fakes.
- Independent read-only review found no substantiated defects; its 26 focused tests
  passed. This does not replace the maintainer's complete release gates.
- `pnpm peers check`, public-import/package contracts and `git diff --check` passed.
  The real rc.1 CLI temporary-home link install/repeat/status/dump/uninstall cycle
  passed without booting a server or touching a live profile.
- Authorized original-instance replay of the formerly failing native-checkpoint plus
  screenshot history returned the requested text with `stop` / `completed` and zero
  tool events after the user restarted. No old history or images were rewritten.
  This proves ordinary replay, not lossless recall or an additional live mixed Basic
  compaction. The accepted host still uses its frozen local R2 link.
- Fixed-tag identity, packaging hashes, and post-publication tag installation results
  are recorded in the GitHub release; local-link acceptance does not prove them.

中文摘要：发布候选通过 545 项测试（另有 1 项既有可选实时测试跳过），并使用正式账户
5.1.3 的精确源码配对验证。原失败会话的真实普通回放已恢复；混合 Basic 压缩为隔离
集成验证，不夸大为新的实机压缩验收。生产仍保留冻结 R2，正式 tag 换装另行安排。

## 0.3.1 + 5.1.2 acceptance / 历史真实验收

The recovery fix has passed an authorized real run on the original instance. The release
pair is compaction `0.3.1` + account `5.1.2`; tag identity is recorded in the GitHub releases.
Release-tag installation and final packaging checks are separate release steps, not inferred
from this run; the frozen production candidate's full regression has also been rerun. See [RECOVERY_FIX.md](RECOVERY_FIX.md) for the behavior and stage history.

- Selected native budget: **300000ms**; elapsed: **157372ms**; **one request**.
- A valid native compaction item and completed event were observed. Official
  `BasicCompactionEngine` produced a **new history replacement**; approximately **146849
  tokens** were shadowed. This is a host history observation, not provider billing or
  evidence of lossless semantic recall.
- The maintainer read `compaction/summary`, the replacement `user/message`, `compaction/end`
  and `command/done` success back from the **disk journal**. This corroborates the recorded
  result for this run, beyond the status command's logical observation; it does not prove
  fsync, crash recovery or elimination of all timeouts.
- The live launcher was `0.1.2-alpha.3`, but the actual Web/Basic dependencies were
  `0.1.2-rc.1`. Do not interpret this mixed environment as full pure-alpha.3 runtime
  compatibility. Public compaction support remains exactly rc.1; the account installer
  retains its existing alpha.3 + rc.1 scope, with both temporary-home install/dump checks
  assigned to final release validation.
- The maintainer reran the frozen production candidate: **498 = plugin 135 + legacy-A 46 +
  comparison 34 + account 266 + paired 17**, all passing. `npm run check` covers the first
  three suites; `scripts/test-accounts.js` covers the isolated account and paired suites.
- `scripts/validate-cli.js` passed the real rc.1 temporary-home install/repeat/status/
  `--dump-config`/uninstall cycle, with no host boot or live-instance operation. Final
  documentation/package/diff checks and account installer checks on both supported CLI
  versions remain distinct release gates.

本次修复已在原实例通过获准的真实验收：新预算 **300000ms**，耗时 **157372ms**，
**一次请求**、合法 item + completed，官方 Basic 产生**新历史替换**，约 **146849 tokens**
被 shadowed。维护者已从**磁盘 journal** 读回 summary、替换 user-message、end 与
command/done success；这是该次记录结果的证据，不证明 fsync、崩溃恢复、无损回忆或全部超时根治。
启动器 alpha.3 与实际 Web/Basic rc.1 的混合现场，不证明纯 alpha.3 完整运行兼容。
维护者已复验冻结生产候选，全量 **498 = 135 + 46 + 34 + 266 + 17** 通过；compaction 的真实
rc.1 临时 home 安装/重复安装/status/dump/卸载闭环通过，不 boot、不碰现网。最终文档/打包/
差异检查、账户安装器两个支持 CLI 版本的临时 home 验证、发布 tag 核验及换装仍是独立门禁。

### Known non-blocking limitation / 已知非阻断限制

Cancellation may still display `CODEX_RUNTIME_ERROR`. Refreshing can cancel a pending manual
command; tab switching alone has not been demonstrated to cause cancellation. This display
issue is retained as a limitation and does not block this release. Persistent upstream failures,
expired budgets and hard context limits remain possible; Native-to-text fallback is this
plugin's policy, not an official Codex fallback guarantee.

取消仍可能误显示 `CODEX_RUNTIME_ERROR`，明确作为不阻断本次发版的限制。刷新可能取消等待中的
手动命令，不证明仅切 tab 就取消。持续上游故障、预算耗尽与窗口硬限制仍可能导致失败；
Native→文本 fallback 是本插件策略，不能称为 Codex 官方保证。

## Historical 0.3.0 environment / 历史 0.3.0 环境

- DSH host: `0.1.2-rc.1`; Node 24.18; macOS arm64.
- SDK pins: auth `@earendil-works/pi-ai` 0.82.1, native protocol module 0.84.4 alias.

## Historical 0.3.0 offline suites / 历史 0.3.0 离线测试

All synthetic data, fake auth/transport, temporary `DSH_HOME`:

- Compaction core: 121 tests; frozen legacy-A baseline: 46; comparison suite: 34.
- Companion account package: 240 tests.
- Paired cross-plugin isolation suite: 12 tests (temporary snapshot install,
  real official basic engine + seam + owner runtime, fake transport).

全部使用合成数据、fake 认证/传输与临时 `DSH_HOME`：压缩核心 121；冻结 legacy-A
基线 46；对比套件 34；配套账户包 240；配对跨插件隔离套件 12（临时快照安装、真实
官方 basic 引擎 + 接缝 + owner 运行时、fake 传输）。

## Historical 0.3.0 controlled live budget / 历史 0.3.0 受控真实预算

A human-approved live trial used 16 real requests: 10 ordinary, 5 native
compactions, 1 stock text compaction; 0 automatic title requests.

经人类批准的真实试用共 16 次请求：10 次普通调用、5 次原生压缩、1 次官方文本压缩；
0 次自动标题请求。

- **Astra (custom model via the trusted metadata seam):** a real automatic
  native replacement, a real ordinary continuation, and a second real native
  compaction ran; after an actual host restart the session recalled the
  synthetic state key exactly (the pre-restart replay plaintext contained no key).
- **Sol (pinned catalog model):** the 5.1.0-rc.1 schema-materialized empty
  input-array defect was fixed in 5.1.0-rc.2; afterwards a real automatic
  native replacement and a real ordinary continuation succeeded. However, the
  model failed to recall tool-borne fixture facts (a synthetic state key and
  three constraints). The official selection verifiably contained those facts;
  an offline bounded simulation showed the local conversion chain and replay
  do not drop them; the actual HTTP exchange was not captured, so no cause is
  concluded here, and the different input paths do not license comparing the
  two models' quality.
- **Legacy B reader:** on the real 5.1.0-rc.2 host, pre-existing structured
  histories passed the current owner binding validation and historical
  checkpoint bytes remained identical (hash values withheld as private).

- **Astra（经可信 metadata 接缝的自定义模型）**：真实自动原生替换、真实普通续写与
  第二次真实原生压缩均完成；在实际宿主重启后，会话精确回忆出合成状态码（重启前
  的重放明文不含该码）。
- **Sol（pinned 目录模型）**：5.1.0-rc.1 的 schema 物化空 input 数组缺陷已在
  5.1.0-rc.2 修复；此后真实自动原生替换与真实普通续写通过。但模型未能回忆工具
  承载的夹具事实（一个合成状态码与三条约束）。官方选区经核验包含这些事实；离线
  有界模拟显示本地转换链与重放不会丢弃它们；实际 HTTP 交互未被捕获，因此此处
  不下任何结论，两条不同的输入路径也不得用于比较两个模型的质量优劣。
- **Legacy B reader**：在真实 5.1.0-rc.2 宿主上，既有结构化历史通过当前 owner
  binding 校验，历史 checkpoint 字节保持相同（哈希值作为私密信息不公开）。

## Release posture / 发布姿态

The profile native default remains **off**; stock basic automatic compaction
stays exactly as-is, and native automation applies only after an explicit
per-session `/codex-native on` (or a separately reviewed profile default).
Release candidates `0.3.0-rc.1` (compaction) and `5.1.0-rc.1`/`5.1.0-rc.2`
(account) are retained as historical tags.

profile 级原生默认保持 **off**；官方 basic 自动压缩原样不变；原生自动化仅在显式
按会话 `/codex-native on`（或另行评审的 profile 默认）之后生效。`0.3.0-rc.1`
（压缩包）与 `5.1.0-rc.1`/`5.1.0-rc.2`（账户包）候选 tag 作为历史保留。