2026-06-06 v1.7.0 - Architecture Progress Alignment and Mainline Plan¶

English Document¶

Objective¶

This document records the concrete mainline plan after re-reading the current code, the active progress dashboards, and the prior architecture plans. The initial 2026-06-06 slice was documentation and governance only; subsequent P1 updates in this same note record release-evidence tooling progress without changing public runtime APIs.

The goal is to make the current main truth explicit:

which parts of the previous RAG / agent / export plans are already code-backed,
which claims are still only operational baselines and not production closure,
where architecture pressure is now concentrated,
what the next forward-compatible and robust implementation sequence should be.

Evidence Base¶

Current branch and workspace state used for this alignment:

branch: main
upstream sync: the initial alignment started from a synced baseline; later local P1 continuation commits may leave main ahead of origin/main until an explicit push
initial worktree state: clean
source files and route registries were re-read from the current workspace, not inferred from older plans

Primary code evidence:

Area	Code evidence	Current reading
Route registry	`src/routes/index.ts`, `src/server.ts`, `src/routes/registry.contract.test.ts`	Modular route registration exists across knowledge, NoteMD, markdown, render, settings, diagnostics, data, and agent-workspace diagnostics. `server.ts` still owns substantial inline orchestration and fallback behavior.
Learning API contract	`src/learning/api.ts`, `src/learning/types.ts`	Knowledge ingest/query/conversation, tutor, memory, quality, guardrail, runtime-state, and workspace-export contracts are typed and exposed through `KnowledgeLearningPlatformAPI`.
Scope and retrieval	`src/learning/types.ts`, `src/learning/KnowledgeLearningPlatform.ts`	`KnowledgeQueryRequest.scope`, `KnowledgeCorpusScope`, workspace readiness, miss diagnostics, planner fields, and scoped conversation traces exist in code.
Conversation runtime	`src/learning/types.ts`, `src/learning/KnowledgeLearningPlatform.ts`, `src/server.ts`, `src/frontend/agent_workspace.js`, `src/frontend/workspace_panes.js`	`AgentConversationRequest`, `AgentConversationResponse`, citations, memory actions, `assistantBlocks`, turn events, stream/replay/turn-cache diagnostics, and operator cards are present.
Durable substrate	`src/resources/`, `src/indexing/`, `src/workspace/`, `src/session/`, `src/workflows/`, `src/memory/`, `src/export/`	Program A-F substrate exists: resources, units/segments, workspace bindings, session/workflow state, governed memory, and deterministic workspace export bundles.
Platform/export profile	`src/platform/PlatformCapabilities.ts`, `src/platform/RenderMaterializer.ts`, `src/routes/render.ts`, `src/export/WorkspaceExportBundle.ts`	Export profiles and PNG-first Godot/mobile materialization are first-class, preserving the Godot SVG limitation boundary.
Rollout governance	`src/learning/store.ts`, `src/learning/queryBackend.ts`, `src/learning/vectorAccelerationAdapter.ts`, `src/learning/runtimeCapability.ts`, `src/server.ts`	Graphdb adapter selection, external HTTP connector telemetry, vector `fail_open` / `fail_closed`, rollout profile payloads, and runbook checks exist.
Architecture pressure	line-count scan on current `main`	`src/server.ts` is about 15,920 lines, `KnowledgeLearningPlatform.ts` about 10,351 lines, `path_app.js` about 4,943 lines, `app.js` about 5,953 lines, and `agent_workspace.js` about 3,237 lines.

Primary plan evidence:

docs/brainstorms/2026-05-25-multiplatform-lightweight-rag-agent-architecture-plan.md
docs/brainstorms/2026-05-26-deep-student-comparison-next-phase-plan.md
docs/solutions/implementation-plan-2026-05-08.md
docs/solutions/implementation-gap-analysis-2026-05-04.md
docs/diataxis/en/explanation/development-progress-dashboard.md
docs/diataxis/zh/explanation/development-progress-dashboard.md

Plan Requirements vs Current Code¶

Prior requirement	Current code reality	Status	Main risk
One canonical conversation contract and reduced API drift	Typed request/response, stream events, replay semantics, turn-cache diagnostics, and frontend contract parity exist. However, the conversation runtime still has major ownership inside `src/server.ts`, so route ownership is not fully converged.	Operational, not fully simplified	Contract growth continues inside the server monolith.
Scope-bound retrieval	`KnowledgeCorpusScope`, workspace/corpus fields, active-target hydration, workspace readiness, and miss diagnostics exist. Program F also supplies resource/index/workspace substrate.	Largely implemented	Release claims still depend on graphdb/ANN calibration, not just scope fields.
Multilingual and lightweight local RAG direction	Local vector and external connector governance exist, with rollout context and ANN runbook gates.	Operational baseline	Recall/latency thresholds and release-grade calibration remain open.
Grounded answer synthesis	`AgentConversationResponse` now includes `answer`, citations, knowledge points, memory actions, trace, and optional `assistantBlocks`. Tauri reply rendering consumes typed blocks when present and falls back to `assistantMessage`.	Operational baseline	More endpoints need to emit richer block payloads without breaking legacy clients.
Memory and invocation loop	Conversation sessions, turns, invocation records, conversation-memory APIs, memory governance, and export inclusion exist.	Implemented baseline	Memory policy is useful but still lighter than a full production memory subsystem with deeper extraction/reranking/audits.
Resource / projection / unit / segment substrate	Program A-F directories and tests exist. Workspace export bundles are deterministic and workspace-scoped.	Implemented	Future work must avoid bypassing this substrate with ad hoc runtime outputs.
Platform shell separation	Export profiles, materialization decisions, PNG-first Godot/mobile behavior, and sidecar/build contracts are explicit.	Implemented	New render/runtime features must keep Godot constraints at adapter boundaries, not in core RAG logic.
Packaging matrix discipline	Tauri desktop, Tauri Android, sidecar readiness, LFS policy, SBOM, and release workflow gates exist.	Strong governance baseline	Some proof is host-dependent; Linux strict Tauri evidence and FR-009 real-device evidence remain release gates.
Architecture reduction	Modular routes and extracted domain modules exist, but `server.ts`, `KnowledgeLearningPlatform.ts`, and frontend host files remain oversized.	Behind target	Adding more orchestration without ownership cuts will worsen maintainability.

Architecture Progress Map¶

Layer	Current maturity	Evidence	Next movement
Graph/path core	Mature operational baseline	`src/core/`, `src/backend/`, path bridge tests	Keep compatibility and focus only on targeted defects.
Runtime storage/retrieval	Operational baseline	`graphdb/sqlite`, external graphdb HTTP, vector acceleration rollout controls	Close soak, latency, recall, and strict rollout calibration before calling production closure.
Scoped RAG/conversation	Operational baseline	scoped query, citations, answer, `assistantBlocks`, turn stream/replay	Move ownership toward a smaller conversation module while preserving response compatibility.
Memory/session/workflow	Implemented substrate	`src/memory/`, `src/session/`, `src/workflows/`, conversation records	Harden policy quality and auditability rather than adding UI-only state.
Export/platform shell	Implemented baseline	export profiles, render materializer, workspace export bundle	Keep mobile/Godot export constraints explicit and test-backed.
Governance and CI	Strong but not complete	docs gates, route/agent contracts, runtime verifiers, SBOM/sidecar gates	Keep evidence freshness and host-specific release gates separate from implementation-complete claims.

Follow-Through on Current `main`¶

The current local main has already moved beyond the initial 2026-06-06 documentation-only alignment slice. Subsequent local mainline commits, still preserving public runtime compatibility, added four concrete execution follow-through tracks:

Foundation release evidence
default freshness audit,
strict repeated-evidence audit,
opt-in multi-host audit,
readiness-exposed release aliases and repeated-evidence checks.
Frontend/runtime auth recovery
stale sidecar settings authentication recovery in the frontend runtime bridge path.
Knowledge Workspace grounded-answer UX
compact API health/status strip,
grouped document-level knowledge hits with matchedSpans,
direct evidence-first top-level answers instead of the old repeated “strongest match” template.
Scoped-retrieval recovery and workspace-local scope UX
planner_scope_recovery when title-like hits exist outside the originally selected scope,
in-pane scope switching inside the Knowledge Workspace,
explicit recovered-source visibility when recovery is used.

These do not change the main planning order. They do change the current progress reading:

P1 is no longer only “release evidence exists”; it now includes freshness, strict repeated evidence, and opt-in multi-host audit tooling.
P3/P4 are no longer only “assistant blocks exist”; they now include intent-aware section composition plus knowledge-workspace scope and grouped-hit UX that makes scoped grounded conversation materially usable in Tauri.
P2 and deeper ownership reduction are still the next structural pressure point, because these newer capabilities remain largely orchestrated through src/server.ts, src/frontend/agent_workspace.js, src/frontend/workspace_panes.js, and KnowledgeLearningPlatform.ts.

Concrete Mainline Plan¶

P0: Keep documentation truth synchronized¶

Treat this document and the development-progress dashboard as the active architecture truth.
Keep docs/en|zh/task.md, docs/en|zh/implementation_plan.md, docs/en|zh/TODO.md, README, and interface docs aligned whenever code or release gates change.
Do not mark a capability production-closed unless runtime evidence and release-grade thresholds both exist.

P1: Finish release-grade graphdb and ANN closure¶

Promote the sqlite soak verifier into repeated release evidence, not a one-off host proof.
Keep verify:foundation:sqlite-runtime:release as the release-facing alias for sqlite soak evidence, and keep foundation readiness mandatory checks aligned with both sqlite and ANN release gates.
Add and keep verify:foundation:release-evidence as the lightweight release-evidence freshness check that reads the latest sqlite soak and ANN release-gate reports before operators treat host evidence as current.
Keep verify:foundation:release-evidence backward-compatible with a minimum of 1 fresh release-contract report per component, use verify:foundation:release-evidence:strict / --min-report-count when a release runbook needs repeated evidence, and use verify:foundation:release-evidence:multi-host / --min-host-count when a release window needs host diversity; foundation readiness exposes the strict repeated-evidence audit as foundation_release_evidence_history.
Treat the new multi-host audit as an executable gate for future release windows, not as a claim that current Windows-host evidence already spans multiple hosts.
Treat old stale or non-release historical reports as warnings in the default audit, while counting only fresh reports that still satisfy the current sqlite soak or ANN release-gate contract.
Current Windows-host strict evidence now passes with sqlite 3/3 and ANN 3/3; this closes local repeated evidence, not multi-host calibration or production closure.
Tighten graphdb connector health/budget thresholds with representative workloads.
Use the new ANN release-gate verifier path (verify:foundation:ann-runtime:release) as the structured evidence entry point for startup, ingest, diagnostics, query latency, and targeted-query recall.
Calibrate ANN recall/latency and external connector behavior under full workload matrices before promoting Phase-2 diagnostics to release gates.

P2: Cut ownership pressure out of `server.ts`¶

Move conversation turn cache, alert trend, runbook bridge, and graphdb rollout helper logic behind explicit modules.
Keep HTTP request validation at route edges and keep internal invariants trusted after validation.
Preserve old response fields such as assistantMessage while making typed block output the preferred path.

P3: Split the learning platform implementation by domain ownership¶

Continue the existing domain extraction direction around ingest, query, conversation, mastery, quality, tutor, and memory.
Extract only when the new owner hides real state or enforces a real invariant.
Avoid pass-through facades that only rename KnowledgeLearningPlatform methods.

P4: Harden the agent workspace contract without widening scope blindly¶

Keep capability registry parity tests as the main drift gate.
Add richer assistant block coverage only when a backend endpoint can produce structured payloads deterministically.
Keep stream-first plus sync fallback behavior and turn replay compatibility intact.

P5: Keep platform and export compatibility explicit¶

Preserve PNG-first Godot/mobile materialization.
Keep desktop-full, desktop-reader, godot-path-mode, and mobile-slim semantics in one export-profile model.
Do not let mobile or Godot constraints leak into the core retrieval/synthesis contract.

Forward-Compatibility and Robustness Guardrails¶

Keep additive response evolution: new fields are optional, old fields remain valid.
Validate at HTTP/runtime boundaries once, then let internal code rely on normalized contracts.
Keep fallback semantics explicit: graphdb fallback, vector fail_open / fail_closed, reader frontend-first plus backend PNG fallback, and Tauri/browser sync fallback must remain observable.
Keep operational evidence separate from production closure. A passing host verifier proves a baseline, not a release threshold.
Avoid changing public endpoint names unless a compatibility shim and tests land in the same slice.
Avoid direct SVG dependence in Godot paths.

Current Verification Position¶

The initial 2026-06-06 alignment slice was documentation-only. The current P1 continuation changes verifier tooling, package scripts, tests, readiness mandatory checks, and documentation while preserving public runtime API compatibility.

The appropriate verification gate for this continuation is:

foundation release-evidence contract test,
default foundation release-evidence audit,
migration test suite,
docs map validation,
docs site build,
markdown/Mermaid fence guard,
git diff review,
final clean worktree after commit.

verify:foundation:release-evidence:strict is intentionally a strict repeated-evidence gate, and getFoundationReadiness().mandatoryChecks exposes it as foundation_release_evidence_history. It now passes on the current Windows host after refreshed sqlite and ANN reports, with sqlite 3/3 and ANN 3/3. verify:foundation:release-evidence:multi-host is now available for release windows that require at least two host keys per component. Other hosts or future release windows must regenerate their own report history before relying on the strict or multi-host gates.

Runtime and test gates for future code slices remain:

npm run verify:foundation:sqlite-runtime:soak
npm run verify:foundation:sqlite-runtime:matrix
npm run verify:foundation:ann-runtime:matrix
npm run verify:foundation:ann-runtime:release
npm run verify:foundation:release-evidence
npm run verify:foundation:release-evidence:strict
npm run verify:foundation:release-evidence:multi-host
npm run test:agent-workspace:contracts
npm run test:migration
npm run verify:core-real-machine:clean

中文文档¶

目标¶

本文档在重新阅读当前代码、活跃进度看板和先前架构方案之后，落盘当前 main 主线的具体推进方案。2026-06-06 首个切片是文档与治理切片；同一文档中的后续 P1 更新会记录 release-evidence 工具推进，但不改变公开运行时 API。

目标是明确当前真实状态：

先前 RAG / agent / export 方案中哪些已经有代码支撑；
哪些只能算 operational baseline，还不能称为 production closure；
当前架构压力集中在哪里；
后续应按什么顺序推进，才能保持向前兼容和鲁棒性。

证据基线¶

本次对齐所依据的分支与工作区状态：

分支：main
远端同步：初始对齐从已同步基线开始；后续本地 P1 延续提交可能会让 main 暂时领先 origin/main，直到显式 push
初始工作区：clean
本次从当前工作区重新读取源码与路由注册表，不沿用旧方案中的推断

主要代码证据：

区域	代码证据	当前判断
路由注册	`src/routes/index.ts`、`src/server.ts`、`src/routes/registry.contract.test.ts`	knowledge、NoteMD、markdown、render、settings、diagnostics、data、agent-workspace diagnostics 已模块化注册；但 `server.ts` 仍承载大量内联编排与 fallback 行为。
学习 API 契约	`src/learning/api.ts`、`src/learning/types.ts`	ingest/query/conversation、tutor、memory、quality、guardrail、runtime-state、workspace-export 均有 typed API 契约。
Scope 与检索	`src/learning/types.ts`、`src/learning/KnowledgeLearningPlatform.ts`	`KnowledgeQueryRequest.scope`、`KnowledgeCorpusScope`、workspace readiness、miss diagnostics、planner 字段与 scoped conversation trace 已落地。
会话运行时	`src/learning/types.ts`、`src/learning/KnowledgeLearningPlatform.ts`、`src/server.ts`、`src/frontend/agent_workspace.js`、`src/frontend/workspace_panes.js`	`AgentConversationRequest/Response`、citation、memory action、`assistantBlocks`、turn events、stream/replay/turn-cache diagnostics 和 operator card 均已存在。
持久化底座	`src/resources/`、`src/indexing/`、`src/workspace/`、`src/session/`、`src/workflows/`、`src/memory/`、`src/export/`	Program A-F 已形成资源、unit/segment、workspace binding、session/workflow state、memory governance、workspace export bundle 底座。
平台/导出 profile	`src/platform/PlatformCapabilities.ts`、`src/platform/RenderMaterializer.ts`、`src/routes/render.ts`、`src/export/WorkspaceExportBundle.ts`	export profile 与 Godot/mobile PNG-first materialization 已是一等契约，保留 Godot SVG 限制边界。
Rollout 治理	`src/learning/store.ts`、`src/learning/queryBackend.ts`、`src/learning/vectorAccelerationAdapter.ts`、`src/learning/runtimeCapability.ts`、`src/server.ts`	graphdb adapter selection、external HTTP connector telemetry、vector `fail_open` / `fail_closed`、rollout profile payload 与 runbook checks 已落地。
架构压力	当前 `main` 行数扫描	`src/server.ts` 约 15,920 行，`KnowledgeLearningPlatform.ts` 约 10,351 行，`path_app.js` 约 4,943 行，`app.js` 约 5,953 行，`agent_workspace.js` 约 3,237 行。

主要方案证据：

docs/brainstorms/2026-05-25-multiplatform-lightweight-rag-agent-architecture-plan.md
docs/brainstorms/2026-05-26-deep-student-comparison-next-phase-plan.md
docs/solutions/implementation-plan-2026-05-08.md
docs/solutions/implementation-gap-analysis-2026-05-04.md
docs/diataxis/en/explanation/development-progress-dashboard.md
docs/diataxis/zh/explanation/development-progress-dashboard.md

方案要求与当前代码对比¶

先前要求	当前代码现实	状态	主要风险
建立唯一 conversation contract 并降低 API 漂移	typed request/response、stream event、replay 语义、turn-cache diagnostics、前端 parity 已存在；但 conversation runtime 的主要所有权仍在 `src/server.ts`。	Operational，但尚未完全简化	合同增长继续堆在 server 单体内。
Scope-bound retrieval	`KnowledgeCorpusScope`、workspace/corpus 字段、active-target hydration、workspace readiness、miss diagnostics 已存在，Program F 也提供 resource/index/workspace 底座。	大体已实现	发布级结论仍取决于 graphdb/ANN 校准，而不是只有 scope 字段。
多语言与轻量 local RAG 方向	local vector 与 external connector 治理已存在，包含 rollout context 与 ANN runbook gates。	Operational baseline	recall/latency 阈值与发布级校准仍未闭环。
Grounded answer synthesis	`AgentConversationResponse` 已包含 `answer`、citation、knowledge point、memory action、trace 和可选 `assistantBlocks`；Tauri 有结构化块渲染与 `assistantMessage` fallback。	Operational baseline	未来更多端点需要产出 richer block payload，同时不能破坏旧客户端。
Memory 与 invocation 闭环	conversation session/turn/invocation、conversation-memory API、memory governance 与 export inclusion 已存在。	已有基线	memory policy 仍比完整生产级 memory subsystem 更轻。
Resource / projection / unit / segment 底座	Program A-F 目录与测试已存在；workspace export bundle 是 deterministic 且 workspace-scoped。	已实现	后续功能不能绕开该底座重新写临时 runtime output。
平台壳层分离	export profile、materialization decision、Godot/mobile PNG-first 行为、sidecar/build 契约已显式化。	已实现	新 render/runtime 功能必须把 Godot 约束留在 adapter 边界。
打包矩阵纪律	Tauri desktop、Tauri Android、sidecar readiness、LFS policy、SBOM、release workflow gate 已存在。	强治理基线	部分证据仍依赖宿主；Linux strict Tauri 与 FR-009 真机证据仍是发布门禁。
架构缩减	模块化路由与领域模块已存在，但 `server.ts`、KLP 和前端宿主文件仍过大。	落后于目标	继续堆编排会恶化可维护性。

架构推进图¶

层级	当前成熟度	证据	下一步
图谱 / Path 核心	成熟 operational baseline	`src/core/`、`src/backend/`、PathBridge 测试	保持兼容，只处理明确缺陷。
运行时存储 / 检索	Operational baseline	`graphdb/sqlite`、external graphdb HTTP、vector acceleration rollout controls	完成 soak、latency、recall 与 strict rollout 校准后再称 production closure。
Scoped RAG / conversation	Operational baseline	scoped query、citation、answer、`assistantBlocks`、turn stream/replay	在保持响应兼容的前提下，把所有权迁入更小的 conversation 模块。
Memory / session / workflow	已实现底座	`src/memory/`、`src/session/`、`src/workflows/`、conversation records	强化 policy 质量与 auditability，而不是增加 UI-only 状态。
Export / platform shell	已实现基线	export profile、render materializer、workspace export bundle	保持 mobile/Godot 导出约束显式且有测试。
Governance / CI	强，但未完全闭环	docs gate、route/agent contract、runtime verifier、SBOM/sidecar gate	把证据新鲜度、宿主相关 release gate 与实现完成度分开描述。

主线具体方案¶

P0：持续同步文档真相¶

以本文档和 development progress dashboard 作为当前架构真相。
只要代码或发布门禁改变，就同步 docs/en|zh/task.md、docs/en|zh/implementation_plan.md、docs/en|zh/TODO.md、README 与接口文档。
没有运行时证据与发布级阈值时，不把 capability 标记为 production-closed。

P1：完成 graphdb 与 ANN 的发布级闭环¶

把 sqlite soak verifier 推进为多轮 release evidence，而不是单次主机证明。
将 verify:foundation:sqlite-runtime:release 保持为 sqlite soak 证据的发布侧别名，并让 foundation readiness mandatory checks 持续对齐 sqlite 与 ANN 两条 release gate。
新增并持续保留 verify:foundation:release-evidence，作为轻量 release-evidence 新鲜度校验：在运维把主机证据视为当前有效证据前，先读取最新 sqlite soak 与 ANN release-gate 报告并验证其仍然新鲜且通过。
保持 verify:foundation:release-evidence 向前兼容：默认每个组件只要求 1 份新鲜且满足 release contract 的报告；当发布 runbook 需要 repeated evidence 时，使用 verify:foundation:release-evidence:strict / --min-report-count；当 release window 需要宿主多样性时，使用 verify:foundation:release-evidence:multi-host / --min-host-count；foundation readiness 会通过 foundation_release_evidence_history 暴露严格 repeated-evidence 审计。
把新的 multi-host 审计视为未来 release window 的可执行门禁，而不是当前 Windows 宿主证据已经覆盖多宿主的结论。
默认审计把旧的过期或非 release 历史报告作为 warning 处理；只有仍然新鲜、且满足当前 sqlite soak 或 ANN release-gate contract 的报告才计入有效数量。
当前 Windows 宿主的 strict evidence 已经以 sqlite 3/3 与 ANN 3/3 通过；这只关闭本地 repeated evidence，不关闭多宿主校准或 production closure。
用代表性 workload 收紧 graphdb connector health/budget 阈值。
将新的 ANN release-gate verifier 路径（verify:foundation:ann-runtime:release）作为 startup、ingest、diagnostics、query latency 与 targeted-query recall 的结构化证据入口。
在把 Phase-2 diagnostics 升级为发布门禁之前，基于完整 workload matrix 完成 ANN recall/latency 与 external connector 行为校准。

P2：从 `server.ts` 切出所有权压力¶

将 conversation turn cache、alert trend、runbook bridge、graphdb rollout helper 等逻辑迁入明确模块。
HTTP 层只做边界校验，内部依赖归一化后的合同。
保留 assistantMessage 等旧响应字段，同时把 typed block 输出作为推荐路径。

P3：按领域所有权拆分学习平台实现¶

沿当前 ingest、query、conversation、mastery、quality、tutor、memory 的领域提取方向继续推进。
只有当新 owner 能隐藏真实状态或强制真实不变量时才提取。
不新增只转发 KnowledgeLearningPlatform 方法的空 facade。

P4：加固 Agent Workspace 合同，不盲目扩面¶

继续以 capability registry parity tests 作为漂移门禁。
只有当后端端点能确定性地产生结构化 payload 时，才扩展 richer assistant block 覆盖面。
保持 stream-first + sync fallback 与 turn replay 兼容。

P5：保持平台与导出兼容性显式化¶

保留 Godot/mobile 的 PNG-first materialization。
把 desktop-full、desktop-reader、godot-path-mode、mobile-slim 放在同一 export-profile 模型中维护。
不让 mobile 或 Godot 约束渗入核心 retrieval/synthesis contract。

向前兼容与鲁棒性护栏¶

响应演进保持 additive：新增字段可选，旧字段继续有效。
在 HTTP/runtime 边界完成一次校验，内部只消费归一化合同。
fallback 语义必须显式可观测：graphdb fallback、vector fail_open / fail_closed、Reader frontend-first + backend PNG fallback、Tauri/browser sync fallback 都要可追踪。
operational evidence 与 production closure 分开写：通过主机 verifier 只能证明 baseline，不等于发布阈值。
除非同一切片落地兼容 shim 与测试，不修改公开 endpoint 名称。
Godot 路径继续禁止依赖直接 SVG 导入。

当前验证位置¶

2026-06-06 初始对齐切片仅修改文档。当前 P1 延续切片会修改 verifier tooling、package scripts、测试、readiness mandatory checks 与文档，但保持公开运行时 API 向前兼容。

当前延续切片的合适验证门禁是：

foundation release-evidence 契约测试；
默认 foundation release-evidence 审计；
migration test suite；
Diataxis 映射校验；
文档站点构建；
Markdown/Mermaid fence 护栏；
git diff review；
提交后工作区 clean。

verify:foundation:release-evidence:strict 是有意严格的 repeated-evidence 门禁，getFoundationReadiness().mandatoryChecks 已通过 foundation_release_evidence_history 暴露它。当前 Windows 宿主在刷新 sqlite 与 ANN 报告后已经以 sqlite 3/3 与 ANN 3/3 通过。verify:foundation:release-evidence:multi-host 现在可用于要求每个组件至少覆盖两个 host key 的 release window。其他宿主或未来 release window 仍必须重新生成自己的历史报告后才能依赖严格或多宿主门禁。

后续代码切片仍应执行：

npm run verify:foundation:sqlite-runtime:soak
npm run verify:foundation:sqlite-runtime:matrix
npm run verify:foundation:ann-runtime:matrix
npm run verify:foundation:ann-runtime:release
npm run verify:foundation:release-evidence
npm run verify:foundation:release-evidence:strict
npm run verify:foundation:release-evidence:multi-host
npm run test:agent-workspace:contracts
npm run test:migration
npm run verify:core-real-machine:clean