Support full init reruns with frozen discovery settings

doc/architecture/SPEC/kb_service.md

@@ -817,12 +817,35 @@ infermatrix-copilot kb init REPO --stage feature-discovery --from-existing \
 ```
 
 `--dry-run` 保存目录和 PR 正文供审查；省略它时沿用 init 的发布权限与 PR 流程。
+已有知识库没有本地 skeleton 记录时，目录 PR 合并后可运行
+`kb init REPO --stage modules --from-existing --pin FULL_SHA`，从已合并的仓库
+入口、owner 路由及页面复核结构，再补充导航。随后使用
+`--stage knowledge --from-existing`；已有 modules 记录仍须完成并合并，
+已有 skeleton 记录的状态与合并门禁同样保留。该模式不生成历史阶段记录，
+不绕过发现目录的源码及哈希绑定；省略 `--from-existing` 时保留原阶段顺序。
+重复初始化若遇到已合并旧批次仍保留的发布分支，可在新批次启动前设置
+`KB_INIT_BRANCH_SUFFIX=<batch-slug>`，使用 `kb/init-REPO-STAGE-SUFFIX`；未设置时
+沿用旧分支名及检查点身份。后缀仅允许 1–40 位小写字母、数字和连字符，
+首尾不能为连字符；非法值在模型调用前拒绝。配置的后缀绑定输入身份及
+已准备发布的分支，恢复发布时不能更换；不会删除或覆盖旧批次分支。
 订阅模式明确使用 Zcode GLM‑5.3 提取、Codex 独立评审，不使用 Claude Code。
 默认并发上限为 13，两轮按顺序运行，轮内共享线程池；Zcode 请求沿用共享排期。
 `KB_DISCOVERY_GENERATOR`、`KB_DISCOVERY_JUDGE`、`KB_DISCOVERY_CONCURRENCY`
 覆盖发现阶段配置；提取与评审必须属于不同模型家族。`--budget-usd` 与
 `--unlimited-subscription` 互斥。未报告实际费用时保持未知，记账值不当作发票费用。
 
+`KB_DISCOVERY_PACKET_CHARS` 设置每次提取组合的正文字符预算（整数 24000–192000，
+默认 24000）；仅组合已有完整分片，不删除或截断输入。`KB_DISCOVERY_START_INTERVAL_S`
+设置 Zcode 初始启动间隔（有限数值 1–60 秒，默认 15 秒）；原生 429/1302 仍触发共享
+冷却和自适应限速。较大的包可能仍有候选遗漏，不能据此声称所有功能已发现。
+`ZCODE_REASONING_LEVEL=low|high|max` 设置实际写入 Zcode 会话配置的推理强度，默认 max；
+Zcode 模型角色中的 `effort` 被 transport 忽略，不能代替该设置。网关追踪和原生 attempt
+均记录 `native_reasoning_level`；发现的输入身份、检查点及紧凑报告 `run_config` 同时绑定
+包大小、启动间隔和实际配置的推理强度。改变任何一项需新批次；旧检查点未记录这些
+配置时不能安全恢复，须保留旧档案并使用新的状态目录。
+返回恰好 24 个候选的分片记录 `candidate_limit_reached`，报告列出数量及任务 ID，
+作为候选可能遗漏的线索；本轮不自动拆分重采样，也不把饱和结果称为穷尽。
+
 发现使用完整文件库存和分片，不把前 N 个文件、前 N 行当作全仓库。范围、排除规则、
 文档入口与功能分母按仓库配置；未知语言使用文本和入口线索，并保留未解析关系。
 没有文档不阻断源码轮，正式功能允许 `docs: []`。目录以外的疑似第一方代码列为

doc/architecture/SPEC/playbooks/PLAYBOOKS.md

@@ -63,6 +63,8 @@ success, steps[]`。
 pr-history → harvest-calibration），每阶段一个由人合并的 PR。只由 `kb init REPO --stage S` 运行，不经 `kb serve`，不写服务账本。
 `feature-discovery` 先提取文档功能基线，再遍历固定源码与测试库存；独立评审后的目录 PR
 合并后冻结，下游核对目录哈希与源码版本。新初始化模板启用发现；旧仓库保持兼容。
+`from_existing` 可用于发现、modules 与解释性知识：缺少本地 skeleton 记录时
+核验已合并入口与 owner 页面；已有阶段记录及发现目录仍保留完成和合并门禁。
 默认 GLM‑5.3 提取、Codex 独立评审，共享并发 13；支持 `from_existing`、
 `retry_unfinished`、`budget_usd` 或 `unlimited_subscription`。未知项不增加正式功能计数，
 原生追踪归档在 Git 外。详细契约见 [`../kb_service.md`](../kb_service.md)。

doc/architecture/SPEC/trace_store.md

@@ -1,6 +1,6 @@
 # trace_store.py —— 规范
 
-<!-- verified-against: 2026-10-04 -->
+<!-- verified-against: 2026-10-05 -->
 
 `LOC ~500 · trace/1：Copilot 与 RB 共用的模型调用/工具调用/判定/结果记录 · refactor-status: stable`
 
@@ -18,6 +18,8 @@
   `native-events.jsonl`，保留原生会话、输出、错误与调度等待事件。结束时绑定唯一的
   `model_call` 回执及输出、用量、耗时与状态；journal 不是另一笔调用，不能重复计费。
   中断未完成的 journal 保留输入和已收到的事件，未报告用量及费用保持未知。
+  对 Zcode，网关调用的 model 与 attempt.model 额外记录 `native_reasoning_level`，
+  表示实际写入会话配置的 low/high/max；role.effort 不是该 transport 的有效推理设置。
 
 - `bind_store(store)` / `current_store()`：把一个 store 绑定为当前进程/任务的采集汇（`tools.dispatch` 与
   `LLM.create` 两个 choke point 据此写 `tool_call` / `model_call`，失败调用记 `error` 且无 `outputs`）；

eval/feature-discovery/README.md

@@ -24,6 +24,14 @@ infermatrix-copilot kb init REPO --stage feature-discovery --from-existing \
 `KB_DISCOVERY_GENERATOR`、`KB_DISCOVERY_JUDGE`、`KB_DISCOVERY_CONCURRENCY` 可覆盖。
 `--budget-usd` 与不限订阅模式互斥。未报告实际费用时保持未知。
 
+批量提取可设置 `KB_DISCOVERY_PACKET_CHARS`（24000–192000，默认 24000），完整索引
+分片和读取范围保持不变。`KB_DISCOVERY_START_INTERVAL_S`（1–60 秒，默认 15）只配置
+启动间隔，仍保留 429/1302 冷却。`ZCODE_REASONING_LEVEL=low|high|max` 控制实际原生
+推理配置；模型角色的 effort 后缀不会替代它。三项设置记录在检查点和报告的
+`run_config`；原生追踪另存 `native_reasoning_level`。配置变化或旧检查点缺少这些身份
+字段时使用新状态目录，保留旧轨迹。扩大包不会证明已发现所有功能，也不取消独立评审。
+恰好达到每包 24 个候选上限的任务另列数量和 ID，保留为潜在遗漏线索。
+
 检查点绑定源码、范围、种子目录、索引版本、提示词与模型配置。成功分片、评审和待审修正稿可复用；预算中断不标为完成。竞争启动同一批次不会覆盖检查点。单项最多三次修正；通道故障保持未知，不重做已成功的提取。仅有文档的未知声明不通过重复改写变成实现证明。
 
 紧凑摘要写入 `eval/feature-discovery/`。完整输入、原生输出、工具事件、配置、用量和修正过程留在 Git 外。恢复时核验每个完成分片的原生归档，包括返回空候选的分片。Zcode 保存原生流式事件；Codex 保存 CLI 结束后按原顺序转发的事件。

playbooks/kb-init.yaml

@@ -18,7 +18,7 @@ params:
   pin: "upstream commit to pin (default: the default branch head)"
   pr_count: "pr-history only: merged upstream PRs to replay (default: adapter pr_history_count or 1000)"
   budget_usd: "discovery/pr-history/depth: cumulative spend ceiling; increasing it resumes the same pinned batch"
-  from_existing: "discovery/explanatory knowledge: enrich a merged index and owner routes without local earlier-stage records"
+  from_existing: "discovery/modules/explanatory knowledge: enrich a merged index and owner routes without local skeleton records; existing stage merge gates remain required"
   retry_unfinished: "feature-discovery/knowledge-deepen: retry unfinished work while retaining accepted work and cumulative spend"
   subscription_generator: "explicit subscription backend; generator USD unreported, fees outside stage USD accounting"
   unlimited_subscription: "feature-discovery/knowledge-deepen: authenticated Zcode GLM-5.3 generator and independent Codex subscription judge without fixed accounting ceiling; incompatible with budget_usd"

plugins/infermatrix-copilot/skills/imkbinit/SKILL.md

@@ -87,17 +87,28 @@ The next stage is the first one that isn't done. If an earlier stage's PR is
 open, stop and say it's waiting for the owner's review. If a record says
 `blocked`, report its `problems` and stop.
 
-For an explicit knowledge rerun of an already merged KB with no local
-skeleton/modules records, use `--stage knowledge --from-existing`. The CLI
+For an explicit breadth rerun of an already merged KB with no local skeleton
+record, use `--stage modules --from-existing`. Merge its PR before running
+`--stage knowledge --from-existing`; no skeleton record is manufactured.
+For a knowledge-only rerun with no local skeleton/modules records, use
+`--stage knowledge --from-existing`. The CLI
 checks the merged index and every owner route. Existing stage records still
 keep their review and merge gates; never manufacture records to skip them.
+An enabled or started feature-discovery batch must still finish and its
+catalog PR must merge before publishing modules or knowledge.
 
 For implementation knowledge in an existing KB, use the independent stage
 `knowledge-deepen` (`kb deepen <repo>` is its `--from-existing` alias).
 It is distinct from the ordinary `deepen` stage that generates hot-module rules.
 `kb widen <repo>` similarly selects `knowledge --from-existing`.
 Use a new state directory when the knowledge baseline or source pin changes;
 retain prior native records and accepted blocks as audit evidence.
+If a previous merged batch's publication branch still exists, set
+`KB_INIT_BRANCH_SUFFIX=<batch-slug>` before starting the new batch. It appends
+the slug to the default stage branch without replacing old branches. Use
+1–40 lowercase letters, digits or hyphens, with no leading or trailing hyphen.
+The suffix is part of the checkpoint identity; retain it when resuming or
+publishing that batch, including a publication waiting for PR confirmation.
 
 ## 3. Run it
 
@@ -143,6 +154,15 @@ What each stage opens:
   Zcode GLM-5.3 extracts and independent Codex reviews with shared concurrency 13.
   `KB_DISCOVERY_GENERATOR`, `KB_DISCOVERY_JUDGE` and `KB_DISCOVERY_CONCURRENCY`
   override the stage configuration; discovery rejects same-family review.
+  `KB_DISCOVERY_PACKET_CHARS` groups complete indexed chunks (24000–192000,
+  default 24000); `KB_DISCOVERY_START_INTERVAL_S` sets the initial Zcode spacing
+  (1–60 seconds, default 15) while native rate-limit cooldown stays enabled.
+  `ZCODE_REASONING_LEVEL=low|high|max` controls the effective native reasoning;
+  a Zcode role's effort suffix does not. These settings are frozen in the batch
+  and report, with reasoning also recorded in both trace and attempt archives.
+  Changed settings or old checkpoints without them require a fresh state
+  directory; retain the previous archive. Larger packets do not prove all
+  features were found.
   `--from-existing` checks merged skeleton routes, while `--retry-unfinished`
   resumes supported work. `--budget-usd` controls cumulative accounting;
   `--unlimited-subscription` authenticates both roles and removes that ceiling.

src/infermatrix_copilot/kb_service/cli.py

@@ -146,8 +146,8 @@ def _init_command(args, state_dir: Path) -> int:
             return 2
         params["acceptance_mode"] = args.acceptance_mode
     if args.from_existing:
-        if args.stage not in ("feature-discovery", "knowledge", "knowledge-deepen"):
-            print("--from-existing is for feature-discovery, knowledge or knowledge-deepen only", file=sys.stderr)
+        if args.stage not in ("feature-discovery", "modules", "knowledge", "knowledge-deepen"):
+            print("--from-existing is for feature-discovery, modules, knowledge or knowledge-deepen only", file=sys.stderr)
             return 2
         params["from_existing"] = "true"
     if args.subscription_generator:
@@ -227,7 +227,7 @@ def main(argv: list[str] | None = None) -> int:
     init.add_argument("--pin", help="upstream commit to pin (default: the default branch head)")
     init.add_argument("--suggest-seeds", action="store_true")
     init.add_argument("--from-existing", action="store_true",
-                      help="enrich a merged KB without local skeleton/modules records (discovery and knowledge stages)")
+                      help="enrich a merged KB without local skeleton records (discovery, modules and knowledge stages)")
     init.add_argument("--subscription-generator", action="store_true",
                       help="explicit subscription generator; unreported fees outside stage USD accounting")
     init.add_argument("--unlimited-subscription", action="store_true",