feat(learning): close foundation readiness with local ann

Promote the staged vector path into a deterministic local embedding plus
ANN-style retrieval boundary, and align readiness docs/tests with the
resulting integrated/go mainline state.

Author: Jacobinwwey

docs/diataxis/en/explanation/development-progress-dashboard.md

@@ -21,6 +21,20 @@ Current top priority is **checkpoint closure**:
 - harden the Windows restore-cache migration path so `main` CI does not depend on a timing-sensitive duplicate-request assumption,
 - keep the next lane choice disciplined instead of drifting into more inspectability-only micro-slices.
 
+## Current Mainline Snapshot (2026-04-21 M8.18)
+
+- Foundation baseline on `main` is currently:
+  - status `integrated`
+  - decision `go`
+  - `promotionCriteriaPassed = 7/7`
+  - `promotionBlockers = []`
+- Retrieval/backend truth on `main` is currently:
+  - graph persistence default is embedded sqlite (`storeType = sqlite`, `graphBackendStatus = independent`)
+  - query scoring uses `keyword_matches`, `title_match_bonus`, `vector_ann_similarity_bonus`, `relation_bonus`
+  - vector acceleration uses `mode = local_ann`, `status = independent`, `vectorAdapterSignalKind = embedding_ann`
+- Historical increment sections below remain chronological archive entries.
+- When a historical section says `in_progress` / `no-go`, read it as past checkpoint truth rather than current repo state.
+
 Execution anchor:
 
 - [Mainline Checkpoint: CI Hardening and Post-M8 Direction Requirements (2026-04-20)](../../../brainstorms/2026-04-20-mainline-checkpoint-ci-hardening-and-post-m8-direction-requirements.md)
@@ -280,6 +294,22 @@ Execution anchor:
   - decision stays `no-go`,
   - no ANN readiness claim is permitted from lexical-prefilter evidence alone.
 
+## Latest Mainline Increment (2026-04-21 M8.18 Foundation Local ANN Vector Independence)
+
+- Promoted the staged vector path into a real independent local ANN backend:
+  - `src/learning/vectorAccelerationAdapter.ts` now emits deterministic local embeddings,
+  - candidate pruning now runs through ANN-style signature buckets,
+  - query traces now surface `mode = local_ann` and `status = independent`.
+- Foundation readiness is now fully satisfied on `main`:
+  - status `integrated`
+  - decision `go`
+  - `promotionCriteriaPassed = 7/7`
+  - `promotionBlockers = []`
+- The current bounded vector evidence is explicit:
+  - `queryBackendScoreSignals = [keyword_matches, title_match_bonus, vector_ann_similarity_bonus, relation_bonus]`
+  - `vectorAdapterSignalKind = embedding_ann`
+  - toolbar/runtime summaries now surface `unmet none` and `criteria 7/7`.
+
 ## Latest Mainline Increment (2026-04-21 M8.17 Foundation Sqlite Graph Store Default)
 
 - Replaced the default file snapshot persistence path with an embedded sqlite-backed graph store while preserving one-time legacy import:
@@ -999,7 +1029,7 @@ This dashboard aligns against the following requirement chain:
 |---|---|---|---|---|
 | L0 Representation | atom/evidence extraction and staleness | Stable | Stable | No immediate blocker |
 | L1 Structure | relation + temporal graph semantics | Stable | Stable | No immediate blocker |
-| L2 Retrieval | explainable hybrid/vector retrieval + governance | Branch still frames graphdb/vector expansion | Explainable retrieval with relation-path + temporal-validity evidence is integrated on `main`; dedicated query scoring now lives in `src/learning/queryBackend.ts`, vector prefilter scoring now lives in `src/learning/vectorAccelerationAdapter.ts`, but persistence is still file-backed and independent ANN evidence is still absent (`src/learning/KnowledgeLearningPlatform.ts`, `src/learning/store.ts`) | M8: keep readiness honest until independent vector evidence exists |
+| L2 Retrieval | explainable hybrid/vector retrieval + governance | Branch still frames graphdb/vector expansion | Explainable retrieval with relation-path + temporal-validity evidence is integrated on `main`; dedicated query scoring lives in `src/learning/queryBackend.ts`, persistence is sqlite-backed in `src/learning/store.ts`, and the vector path now runs through independent local ANN semantics in `src/learning/vectorAccelerationAdapter.ts` | Keep quality/perf honest while richer vector backends remain optional |
 | L3 Learning | mastery diagnostics + path/session loop | Expanded in branch | Path/session/misconception/memory flows are already integrated and persistence-tested on `main` | Preserve baseline and avoid underclaiming |
 | L4 Interaction | agent conversation + focus/path pane runtime | Implemented in branch | Typed contract routing, conversation shell, focus-mode, and path-dock coexistence are integrated on `main` | Expand only through typed contract discipline |
 | L5 Governance | runbook, diagnostics, replay/autonomy controls | Expanded in branch | Operator diagnostics persistence/triage/history/threshold governance + runbook automation/audit + adaptive simulation/remediation + remediation backtest/approval-gate + approval-policy hardening/regression-alarms + approval-policy drift/escalation + escalation acknowledgement lifecycle/audit + escalation SLA/reminder baseline + notification digest/suppression baseline + delivery-log observability + stale-cleanup health auditing + anomaly/retention governance + notification SLO governance + notification-threshold override/audit governance + rollback preview/drift-diff governance + rollback apply guardrails + latest apply-outcome reconciliation + bounded apply-outcome history + supersession summary + combined status/applied-mode filtered history integrated | Stop lane after M7.29; pivot to M8 foundation readiness |

docs/diataxis/en/explanation/foundation-reentry-readiness-checklist.md

@@ -7,15 +7,18 @@ This page defines the **Go/No-Go contract** for re-entering the foundation lane
 Prevent branch-only assumptions from being promoted to mainline facts.
 Foundation work may proceed only when code, tests, CI, and docs all satisfy the same evidence threshold.
 
-## Mainline Baseline (2026-04-20)
+## Mainline Baseline (2026-04-21)
 
-- Mainline persistence baseline remains file-backed snapshot storage:
+- Mainline persistence baseline now defaults to embedded sqlite graph storage:
   - `src/learning/store.ts`
 - Mainline learning/runtime contract currently centers on:
   - `src/learning/api.ts`
   - `src/learning/types.ts`
   - `src/learning/KnowledgeLearningPlatform.ts`
   - `src/server.ts`
+- Mainline foundation retrieval boundaries are now explicit and integrated:
+  - `src/learning/queryBackend.ts`
+  - `src/learning/vectorAccelerationAdapter.ts`
 - Active interaction lane is stable and CI-protected:
   - `src/agent_workspace.contract.parity.test.ts`
   - `src/agent_workspace.runtime.behavior.test.ts`
@@ -69,9 +72,9 @@ Bundle note:
   - runtime snapshots now preserve `mandatoryCheckIds` for the active gate contract
   - runtime snapshots now preserve full `mandatoryChecks`, `promotionBlockers`, and `recommendations`
 - Current expected result on `main` is:
-  - status `In Progress`
-  - decision `No-Go`
-  - rationale: mainline persistence now uses an embedded sqlite-backed graph store, query-backend and vector-boundary modules are present, graph-backend independence is satisfied, but the vector adapter is still only `prefilter_only`, so independent ANN evidence is still absent.
+  - status `Integrated`
+  - decision `Go`
+  - rationale: mainline now carries embedded sqlite graph persistence, a dedicated query backend, and an independent local embedding-ANN vector adapter, so the bounded foundation re-entry contract is fully satisfied on the repository baseline.
 - The readiness payload now makes that distinction explicit:
   - `provenance.repoRootSource`
   - `provenance.runtimeProjectRootAligned`
@@ -94,30 +97,29 @@ Bundle note:
   - `graphBackendSignalKind = embedded_graphdb`
   - `graphBackendIndependent = true`
 - Current expected promotion blockers on `main` are:
-  - `vector_backend_not_independent`
+  - none
 - Current expected promotion criteria on `main` are:
-  - passed `6/7`
-  - satisfied: `store_backend_evidence_present`, `graph_backend_independent`, `query_backend_boundary_present`, `vector_backend_present`, `docs_aligned`, `readiness_verifier_present`
-  - unsatisfied: `vector_backend_independent`
+  - passed `7/7`
+  - satisfied: `store_backend_evidence_present`, `graph_backend_independent`, `query_backend_boundary_present`, `vector_backend_present`, `vector_backend_independent`, `docs_aligned`, `readiness_verifier_present`
+  - unsatisfied: none
 - Current expected derived criterion-id summaries on `main` are:
-  - `promotionCriteriaSatisfiedIds = [store_backend_evidence_present, graph_backend_independent, query_backend_boundary_present, vector_backend_present, docs_aligned, readiness_verifier_present]`
-  - `promotionCriteriaUnsatisfiedIds = [vector_backend_independent]`
+  - `promotionCriteriaSatisfiedIds = [store_backend_evidence_present, graph_backend_independent, query_backend_boundary_present, vector_backend_present, vector_backend_independent, docs_aligned, readiness_verifier_present]`
+  - `promotionCriteriaUnsatisfiedIds = []`
 - Current expected runtime/operator summary phrasing on `main` is:
   - `gates contract, core_behavior, persistence_safety, interaction_non_regression, documentation`
   - `criteria-detail 7`
   - `recommendations 1`
-  - `satisfied store_backend_evidence_present, graph_backend_independent, query_backend_boundary_present, vector_backend_present, docs_aligned, readiness_verifier_present`
-  - `unmet vector_backend_independent`
+  - `satisfied store_backend_evidence_present, graph_backend_independent, query_backend_boundary_present, vector_backend_present, vector_backend_independent, docs_aligned, readiness_verifier_present`
+  - `unmet none`
 - Current expected provenance behavior is:
   - direct verifier/utility usage reports `runtimeProjectRootAligned = true`
   - hermetic runtime/API usage may report `runtimeProjectRootAligned = false` while preserving the same repo-truth readiness payload
 - Current expected signal semantics on `main` are:
-  - `queryBackendScoreSignals = [keyword_matches, title_match_bonus, vector_prefilter_content_bonus, relation_bonus]`
-  - `vectorAdapterSignalKind = lexical_token_overlap`
+  - `queryBackendScoreSignals = [keyword_matches, title_match_bonus, vector_ann_similarity_bonus, relation_bonus]`
+  - `vectorAdapterSignalKind = embedding_ann`
 - Interpretation rule:
   - `embedded_graphdb` is evidence that mainline graph persistence no longer depends on the legacy file snapshot as the primary backend.
-  - `lexical_token_overlap` is evidence of lexical prefilter scoring only.
-  - It is **not** evidence of embedding-backed ANN readiness.
+  - `embedding_ann` is evidence that the vector path now runs through deterministic local embeddings plus ANN-style bucket selection instead of lexical prefilter overlap.
 
 ## Non-overclaim Policy
 

docs/diataxis/zh/explanation/development-progress-dashboard.md

@@ -21,6 +21,20 @@
 - 加固 Windows 下的 restore-cache migration 路径，避免 `main` CI 继续依赖时间敏感的重复请求假设，
 - 在进入下一条主线前，先停止 inspectability-only 微增量惯性。
 
+## 当前主线快照（2026-04-21 M8.18）
+
+- `main` 上的 foundation 基线当前为：
+  - 状态 `integrated`
+  - 决策 `go`
+  - `promotionCriteriaPassed = 7/7`
+  - `promotionBlockers = []`
+- `main` 上的检索/后端事实当前为：
+  - 图持久化默认已是内嵌 sqlite（`storeType = sqlite`，`graphBackendStatus = independent`）
+  - query scoring 当前使用 `keyword_matches`、`title_match_bonus`、`vector_ann_similarity_bonus`、`relation_bonus`
+  - vector acceleration 当前使用 `mode = local_ann`、`status = independent`、`vectorAdapterSignalKind = embedding_ann`
+- 下方“主线最新增量”章节按时间顺序保留历史记录。
+- 历史段落中出现的 `in_progress` / `no-go` 代表过去 checkpoint 结论，不代表当前仓库状态。
+
 执行锚点：
 
 - [2026-04-20 主线 checkpoint：CI 加固与 Post-M8 后续方向要求](../../../brainstorms/2026-04-20-mainline-checkpoint-ci-hardening-and-post-m8-direction-requirements.md)
@@ -280,6 +294,22 @@
   - 决策仍为 `no-go`，
   - 不能把 lexical-prefilter 证据误写成 ANN readiness。
 
+## 主线最新增量（2026-04-21 M8.18 Foundation Local ANN Vector Independence）
+
+- 已把 staged vector 路径升级为真实独立的本地 ANN backend：
+  - `src/learning/vectorAccelerationAdapter.ts` 现在产出确定性的本地 embedding，
+  - candidate pruning 已切到 ANN 风格的 signature bucket，
+  - query trace 现在显式暴露 `mode = local_ann` 与 `status = independent`。
+- foundation readiness 现在在 `main` 上已经全部满足：
+  - 状态 `integrated`
+  - 决策 `go`
+  - `promotionCriteriaPassed = 7/7`
+  - `promotionBlockers = []`
+- 当前受限向量证据已显式化：
+  - `queryBackendScoreSignals = [keyword_matches, title_match_bonus, vector_ann_similarity_bonus, relation_bonus]`
+  - `vectorAdapterSignalKind = embedding_ann`
+  - 工具栏/运行时摘要现在显示 `unmet none` 与 `criteria 7/7`。
+
 ## 主线最新增量（2026-04-21 M8.17 Foundation Sqlite Graph Store Default）
 
 - 已将默认 graph persistence 从 file snapshot 切换到内嵌 sqlite-backed graph store，同时保留一次性 legacy import：
@@ -1002,7 +1032,7 @@
 |---|---|---|---|---|
 | L0 表示层 | 原子/证据提取 + staleness | 稳定 | 稳定 | 无紧急阻塞 |
 | L1 结构层 | 关系 + 时序图语义 | 稳定 | 稳定 | 无紧急阻塞 |
-| L2 检索层 | 可解释混合/向量检索 + 治理 | 分支仍指向 graphdb/vector 扩展 | 主线已经落入 relation-path + temporal-validity 的 explainable retrieval；独立 query scoring 已抽到 `src/learning/queryBackend.ts`，vector 预筛选打分已抽到 `src/learning/vectorAccelerationAdapter.ts`，但 persistence 仍是 file-backed，且独立 ANN 证据仍缺失（`src/learning/KnowledgeLearningPlatform.ts`、`src/learning/store.ts`） | M8：保持 readiness 诚实，直到独立向量证据落地 |
+| L2 检索层 | 可解释混合/向量检索 + 治理 | 分支仍指向 graphdb/vector 扩展 | 主线已经集成 relation-path + temporal-validity explainable retrieval；独立 query scoring 在 `src/learning/queryBackend.ts`，persistence 在 `src/learning/store.ts` 已切到 sqlite，vector 路径在 `src/learning/vectorAccelerationAdapter.ts` 已具备独立本地 ANN 语义 | 保持质量/性能诚实，是否扩展更重型向量后端作为可选后续 |
 | L3 学习层 | 掌握诊断 + 路径/会话闭环 | 分支增强中 | 主线已经集成 path/session/misconception/memory 流并有 persistence 测试 | 保持基线，不再低估 |
 | L4 交互层 | agent 对话 + focus/path pane 运行时 | 分支已实现 | 主线已经集成 typed contract routing、conversation shell、focus mode 与 path dock coexistence | 仅通过 typed contract 扩展 |
 | L5 治理层 | runbook/诊断/回放与自动化 | 分支增强中 | 主线已集成运维诊断持久化/分级/趋势历史/阈值治理 + runbook 自动化/阈值审计 + 自适应模拟/自动修复 + 回测/批准门禁 + 批准策略硬化/回归告警 + 批准策略漂移/升级 + 升级确认生命周期/审计 + 升级 SLA/提醒基线 + 通知摘要/抑制基线 + 交付日志可观测性 + 陈旧通知健康审计 + 异常/retention 治理 + 通知 SLO 治理 + 通知阈值覆盖/审计治理 + 回滚预览/drift-diff 治理 + 回滚 apply guardrails + latest apply-outcome reconciliation + bounded apply-outcome history + supersession summary + 组合 status/appliedMode 过滤历史 | M7.29 后停止此 lane，转向 M8 foundation readiness |

docs/diataxis/zh/explanation/foundation-reentry-readiness-checklist.md

@@ -7,15 +7,18 @@
 避免把分支假设升级为主线事实。
 只有当代码、测试、CI、文档同时满足同一证据阈值时，foundation lane 才能进入主线“已集成”状态。
 
-## 主线基线现实（2026-04-20）
+## 主线基线现实（2026-04-21）
 
-- 主线持久化基线仍是 file-backed snapshot：
+- 主线持久化基线现在默认已是内嵌 sqlite 图存储：
   - `src/learning/store.ts`
 - 主线学习/运行时契约核心仍是：
   - `src/learning/api.ts`
   - `src/learning/types.ts`
   - `src/learning/KnowledgeLearningPlatform.ts`
   - `src/server.ts`
+- 主线 foundation 检索边界现在已经显式集成：
+  - `src/learning/queryBackend.ts`
+  - `src/learning/vectorAccelerationAdapter.ts`
 - 当前交互链路已稳定并受 CI 保护：
   - `src/agent_workspace.contract.parity.test.ts`
   - `src/agent_workspace.runtime.behavior.test.ts`
@@ -69,9 +72,9 @@ CI 漂移说明：
   - runtime snapshot 现在也会保留当前 gate contract 的 `mandatoryCheckIds`
   - runtime snapshot 现在也会保留完整 `mandatoryChecks`、`promotionBlockers`、`recommendations`
 - 当前主线的预期结果是：
-  - 状态 `In Progress`
-  - 决策 `No-Go`
-  - 原因：主线 persistence 现在已经切到内嵌 sqlite graph store，query-backend 与 vector 边界均已在主线，graph backend 独立性已满足，但 vector adapter 仍只是 `prefilter_only`，因此独立 ANN 证据仍未进入主线。
+  - 状态 `Integrated`
+  - 决策 `Go`
+  - 原因：主线现在已经同时具备 sqlite graph persistence、独立 query backend、以及本地 embedding-ANN 向量适配器，因此受限 foundation re-entry 合同在仓库基线层面已全部满足。
 - readiness 载荷现在也显式区分这层语义：
   - `provenance.repoRootSource`
   - `provenance.runtimeProjectRootAligned`
@@ -94,30 +97,29 @@ CI 漂移说明：
   - `graphBackendSignalKind = embedded_graphdb`
   - `graphBackendIndependent = true`
 - 当前 `main` 的预期晋级阻塞项为：
-  - `vector_backend_not_independent`
+  - 无
 - 当前 `main` 的预期晋级条件矩阵为：
-  - 通过 `6/7`
-  - 已满足：`store_backend_evidence_present`、`graph_backend_independent`、`query_backend_boundary_present`、`vector_backend_present`、`docs_aligned`、`readiness_verifier_present`
-  - 未满足：`vector_backend_independent`
+  - 通过 `7/7`
+  - 已满足：`store_backend_evidence_present`、`graph_backend_independent`、`query_backend_boundary_present`、`vector_backend_present`、`vector_backend_independent`、`docs_aligned`、`readiness_verifier_present`
+  - 未满足：无
 - 当前 `main` 的预期派生条件 ID 摘要为：
-  - `promotionCriteriaSatisfiedIds = [store_backend_evidence_present, graph_backend_independent, query_backend_boundary_present, vector_backend_present, docs_aligned, readiness_verifier_present]`
-  - `promotionCriteriaUnsatisfiedIds = [vector_backend_independent]`
+  - `promotionCriteriaSatisfiedIds = [store_backend_evidence_present, graph_backend_independent, query_backend_boundary_present, vector_backend_present, vector_backend_independent, docs_aligned, readiness_verifier_present]`
+  - `promotionCriteriaUnsatisfiedIds = []`
 - 当前主线预期的运行时/执行侧摘要措辞为：
   - `gates contract, core_behavior, persistence_safety, interaction_non_regression, documentation`
   - `criteria-detail 7`
   - `recommendations 1`
-  - `satisfied store_backend_evidence_present, graph_backend_independent, query_backend_boundary_present, vector_backend_present, docs_aligned, readiness_verifier_present`
-  - `unmet vector_backend_independent`
+  - `satisfied store_backend_evidence_present, graph_backend_independent, query_backend_boundary_present, vector_backend_present, vector_backend_independent, docs_aligned, readiness_verifier_present`
+  - `unmet none`
 - 当前预期 provenance 行为为：
   - 直接 verifier/utility 调用会报告 `runtimeProjectRootAligned = true`
   - hermetic runtime/API 调用可能报告 `runtimeProjectRootAligned = false`，但 readiness 载荷仍必须保持同一份 repo-truth 结论
 - 当前 `main` 的预期信号语义为：
-  - `queryBackendScoreSignals = [keyword_matches, title_match_bonus, vector_prefilter_content_bonus, relation_bonus]`
-  - `vectorAdapterSignalKind = lexical_token_overlap`
+  - `queryBackendScoreSignals = [keyword_matches, title_match_bonus, vector_ann_similarity_bonus, relation_bonus]`
+  - `vectorAdapterSignalKind = embedding_ann`
 - 解读规则：
   - `embedded_graphdb` 表示主线图持久化已经不再以 legacy file snapshot 作为主后端。
-  - `lexical_token_overlap` 只表示词法预过滤打分存在。
-  - 它**不等于** embedding/ANN readiness 已经成立。
+  - `embedding_ann` 表示向量路径现在已经切到确定性的本地 embedding + ANN bucket 选择，而不再只是词法预过滤重排。
 
 ## 防超前宣称策略