diff --git a/.gitattributes b/.gitattributes
new file mode 100644
index 00000000..adb9bfa9
--- /dev/null
+++ b/.gitattributes
@@ -0,0 +1,14 @@
+# Line-ending policy.
+#
+# Shell scripts must keep LF in the working tree: a CRLF checkout makes
+# `bash scripts/*.sh` fail with `$'\r': command not found`, and a Windows
+# working copy is also what a local Docker build would copy into the
+# containers, so the breakage is not limited to the developer's own shell.
+#
+# `text eol=lf` overrides core.autocrlf for the matched paths, so the outcome
+# no longer depends on each contributor's local Git configuration
+# (github.com docs, "Configuring Git to handle line endings": a committed
+# .gitattributes "overrides the core.autocrlf setting for all repository
+# contributors"; gitattributes(5): the eol attribute "sets a specific
+# line-ending style to be used in the working directory").
+*.sh text eol=lf
diff --git a/AGENTS.md b/AGENTS.md
index 58eccbb1..1c2d270c 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -156,4 +156,4 @@ snapshot、HypothesisSpec 确定性编译、五代 8×3 共演化、单冠军 Fo
 
 ---
 
-> 本文件是协议入口，**短小**为美。详细规则在公开的 `docs/00-03` 与 `docs/brand/`。
+> 本文件是协议入口，**短小**为美。详细规则在公开的 `docs/00-05`。
diff --git a/agent_onboarding.md b/agent_onboarding.md
new file mode 100644
index 00000000..9951b9f1
--- /dev/null
+++ b/agent_onboarding.md
@@ -0,0 +1,316 @@
+# Agent Onboarding — Inalpha
+
+> Audience: an AI coding agent (or a new human contributor) picking up this repository cold.
+> Derived from a full read-only familiarization pass of HEAD `4822bb9` (2026-10-14).
+> This document is a map and a contract, not a substitute for the code. Where this file and the
+> code disagree, **the code wins** — and fix this file in the same change.
+
+## 1. Working agreements
+
+These are binding for any agent working in this repository.
+
+- Never commit or document sensitive information. No API keys, secrets, private keys, wallet material, tokens, passwords, or personal data in code, tests, fixtures, docs, commit messages, or logs. Adapter credentials stay in environment variables. Read credentials from the environment only, and never add an environment variable to a test in a way that makes the test depend on it.
+
+- Always commit and push code after a code change. A code change is not finished when the file is written; it is finished when the change is committed and pushed.
+
+- Fix defects on the spot, unless the owner decides otherwise. When a defect is found in the code being worked on, fix it as part of the same change rather than reporting it and moving on. Defer only when the owner explicitly decides to defer, and record the deferral.
+
+- Never make code changes unless they are defects. Keep the change surface limited to defect fixes. Do not add unrelated features, refactors, renames, abstractions, drive-by cleanups, or speculative improvements. Note unrelated observations instead of changing them.
+
+- Always keep corresponding documentation in sync with code changes. A code change updates the documents and generated artifacts that describe it, in the same change. This includes concept and integration docs, coding-standards guidance, and the generated stubs and docstrings. Maintainer-owned files such as RELEASES.md are updated by maintainers, not by agents.
+
+- **Always do an extensive and deep web search before deciding, or when the answer is unknown.** When a decision has more than one defensible option, or the answer is not already established by this repository's own code, tests and docs, search the web FIRST — broadly, across several independent sources — and read the PRIMARY source before deciding: the vendor/API documentation, the specification, the issue tracker, or the pinned dependency's own source at the version this repository locks. Confirm that version, because an API that changed after an old forum answer sends you the wrong way, and cross-check any claim the decision rests on against a second independent source. Record what the search established, with its sources, in the change notes (the commit message, or the doc the change touches) whenever it determined the outcome. "Found nothing" is a result: report it, and label the choice an assumption — never present a guess as researched. A search never overrides evidence this repository already holds: for internal behavior the code, tests and stored artifacts are ground truth, and a run you measured outranks anything published.
+
+Repository-level rules that layer on top of the above live in [`AGENTS.md`](AGENTS.md) (hard constraints,
+naming, "don't touch" list, GitHub collaboration language) and [`CONTRIBUTING.md`](CONTRIBUTING.md)
+(branch model, commit conventions, review flow). Read both before your first change.
+
+---
+
+## 2. What this repository is
+
+**Inalpha** is an AI-agent-orchestrated, multi-kernel quant research and paper-trading framework.
+It is a multi-service **application**, not a library: a Next.js operator console drives a Mastra
+(TypeScript) orchestration layer, which drives five independent Python FastAPI services over HTTP.
+
+| Layer | Component | Port | Role |
+|---|---|---|---|
+| Entry | `apps/dashboard` | 3001 | Operator console: auth, same-origin BFF, docked agent chat, portfolio / runners / risk / factors / lab / evolution pages |
+| Entry | `apps/web` | — | Static marketing site (`output: "export"`) |
+| Entry | `mastra dev` playground | 4111 | Live trace / debug surface for the orchestrator |
+| Orchestration | `packages/orchestration` | 4111 | One orchestrator agent, ~60 HTTP tools, hooks, permissions, plan/exec proxies, scheduler, MCP, skills, evals |
+| Kernel | `services/data` | 8001 | Market data, fundamentals, news, web search, FX, point-in-time event ledger |
+| Kernel | `services/paper` | 8002 | Event-driven kernel, backtest, live runner, strategy authoring sandbox, plan/exec, risk, wallets |
+| Kernel | `services/research` | 8003 | Multi-analyst + bull/bear/risk debate → `StrategyHint` |
+| Kernel | `services/factor` | 8004 | Factor library, IC/decay, DSL discovery, panel scoring (signals only) |
+| Kernel | `services/evolver` | 8005 | E1 single-generation mutation; E2 event campaigns + persistent `EvolutionLoop` |
+| Shared | `services/_shared` | — | `inalpha_shared` (JWT, psycopg pool, errors, logging, usage ledger) + `_shared/llm` |
+| Persistence | Postgres 17 + TimescaleDB, Redis | — | All business state; `bars` / `ticks` hypertables; Mastra memory is LibSQL/SQLite |
+
+Status: self-declared **alpha**. Real-money trading is explicitly out of scope — the ceiling is
+simulated ("paper") accounts. `docs/04-current-state.md` is the authoritative status page: it states
+that the E2 engineering validation and the real positive production acceptance are **not** complete.
+
+---
+
+## 3. Ground rules — do not break these
+
+These are enforced in code today; each maps to a test or an explicit raise. Treat violations as defects.
+
+1. **The LLM has no order path.** The only sanctioned route is `trade.create_plan → approve_plan →
+   execute_plan`, with a one-time, TTL-bound `approval_token` minted and consumed by `services/paper`.
+   Direct-submit tool ids are denied *and* absent from the model's tool list.
+2. **Permission resolution is deny → allow → ask → default(`ask`)**, and a layer merge can never
+   downgrade a `deny`. Every `ask` requires a verified owner subject and a stable session id, else
+   fail-closed with no side effects.
+3. **Risk is enforced before execution.** HTTP path: `RiskGuard.enforce` takes its own DB connection
+   and commits the lock *before* raising, so the lock survives the caller's rollback. The automated
+   live runner is fail-closed without a guard factory. Note the asymmetry: if the HTTP risk config
+   fails to load, the manual HTTP path is **fail-open** — do not widen that hole.
+4. **No look-ahead.** Orders submitted on bar N fill on bar N+1 at the next open. Market events are
+   visible only when `available_at <= bar.bar_known_at`. The live runner consumes only closed bars.
+5. **Simulation time never goes backwards** (`TestClock.set_time` / `advance_time` raise).
+6. **Order state machine legality and overfill rejection** are enforced on the domain model, not at
+   the edges.
+7. **No naked short, no cash overdraw, perp initial margin ≤ wallet** — checked before execution *and*
+   re-checked under `FOR UPDATE` inside the fill transaction.
+8. **Per-run wallets are isolated**; lock order is account → wallet → position; one running run per
+   candidate per owner; per-account run cap applies.
+9. **Promotion requires a completed backtest**; starting a live run requires `status = 'promoted'`
+   plus owner match.
+10. **Candidate source passes three gates** — AST audit, restricted loader, `Strategy` contract check —
+    re-run at authoring, at backtest, and at live build. Guard-order forgery by literal string is rejected.
+11. **Evolution is authorized, owner-scoped, fenced and budgeted.** No auto-promote, no auto-start,
+    no auto-order. Adopted artifacts are literally `runner_eligible: false`. E1 needs per-run explicit
+    approval; E2 behind `EVENT_EVOLUTION_ENABLED` authorizes a bounded loop, not generations.
+12. **Secrets never enter run records, logs, or the browser.** Per-user LLM keys are encrypted at rest,
+    resolved per owner at call time, and cleared after the credential exchange.
+13. **Freshness is a correctness property.** Treat "looks new but is stale" as a bug; reason about
+    `bars[-1].ts` distance from `as_of`, never about bar count.
+
+---
+
+## 4. Repository map — where to look for what
+
+| You are asked to change | Primary location | Also affected |
+|---|---|---|
+| A model-visible tool | `packages/orchestration/src/tools/<area>.ts`, bucket in `tools/index.ts` | `packages/orchestration/config/permissions.default.yaml`, `mastra/tool-wiring.ts`, tool-description style |
+| Guardrail policy | `packages/orchestration/config/permissions.default.yaml` (+ `permissions/defaults.ts` mirror) | `hooks/with-hooks.ts`, `permissions/pending.ts`, dashboard approval API |
+| Order / approval flow | `services/paper/src/inalpha_paper/api/trade_plans.py`, `storage/trade_plans.py` | `orchestration/src/tools/trade-plan.ts`, `api/orders.py`, `live_runner.py` |
+| Kernel semantics (bars, fills, clock) | `services/paper/src/inalpha_paper/engine/backtest.py`, `execution/*`, `kernel/*` | `engine/live_session.py`, every strategy, factor/evolver evaluation |
+| Live runner | `services/paper/src/inalpha_paper/live_runner.py` | `api/strategy_runs.py`, `storage/run_wallets.py`, dashboard `/api/runners*` |
+| Risk rules | `services/paper/src/inalpha_paper/execution/risk_rules/*`, `configs/risk_rules.toml` | `risk_guard.py`, `api/orders.py`, `api/trade_plans.py`, `live_runner.py` |
+| Data connectors / freshness | `services/data/src/inalpha_data/connectors/*`, `api/{bars,backfill,ticker}.py` | all consumer clients, orchestrator prompts |
+| Factors | `services/factor/src/inalpha_factor/{adapters,engine,effectiveness}.py` | `paper/factor_patrol.py`, `factor.*` tools, dashboard factors page |
+| Evolution (E1) | `services/evolver/src/inalpha_evolver/runtime/{executor,generation,slots}.py`, `mutator/*` | `paper/strategy_authoring/*` (imported!), `evolver.*` tools |
+| Evolution loops/campaigns (E2) | `runtime/{loop_manager,loop_campaign,campaign}.py`, `loop_fencing.py`, `api/{loop_start,campaign_routes}.py` | feature flag, data event snapshots, paper Forward sandbox |
+| Auth / per-user keys | `services/_shared/src/inalpha_shared/auth.py`, dashboard `lib/{session,encryption,backend}.ts` | `orchestration/src/{auth.ts,mastra/identity.ts}` |
+| DB schema | `infra/migrations/versions/<new>.py` (Alembic, raw `op.execute` SQL) | `storage/*` in the owning service |
+| Console UI | `apps/dashboard/src/app/[locale]/**`, `src/components/**` | `apps/dashboard/design.md`, shared `components/ui/*` |
+| Public docs | `README*.md`, `docs/00-05`, `AGENTS.md`, `CLAUDE.md` | run `bash scripts/check-consistency.sh` |
+
+---
+
+## 5. Execution paths you must understand before editing
+
+**Chat turn to tool call**
+
+```
+browser -> dashboard /api/copilotkit (BFF: decrypt owner LLM key, inject header)
+  -> mastra :4111 -> identity middleware (verify Bearer, bind user LLM config)
+    -> orchestrator -> tool.execute  [wrapped by hooks]
+         PreToolUse hooks -> PermissionEngine.authorize -> [ask: approval store]
+         -> clients/<service>.ts -> HTTP to the Python service
+         -> PostToolUse / PostToolUseFailure hooks
+```
+
+**Order path (the only route to a trading side effect)**
+
+```
+trade.create_plan -> POST /plans (pending, TTL)
+trade.approve_plan -> POST /plans/{id}/approve -> one-time approval_token
+trade.execute_plan -> POST /plans/{id}/execute -> consume token -> RiskGuard.enforce
+                   -> order insert + positions + cash in one transaction
+```
+
+**Live runner bar loop**
+
+```
+POST /strategy_runs -> check promoted + owner + cap + allocation
+  -> background task -> per run: poll closed bars -> dedup by last_bar_ts
+     -> per bar: notional cap -> DB risk enforce -> spot/perp prechecks
+        -> order executor -> single transaction (account lock, status re-check,
+           FOR UPDATE clamps, plan create/approve/consume, order insert,
+           fills -> positions + cash, decision row)
+```
+
+**Backtest**
+
+```
+POST /backtest -> validate -> data client (user Bearer forwarded) -> fresh backfill
+  -> process pool -> evaluate candidate (+ buy-and-hold baseline) -> fitness
+  -> optional holdout block -> persist run + trades in one transaction -> report
+```
+
+---
+
+## 6. Commands
+
+```bash
+# install
+cd packages/orchestration && pnpm i && cd ../..
+for s in data paper research factor evolver; do (cd "services/$s" && uv sync); done
+
+# database
+cp infra/.env.example infra/.env && (cd infra && docker compose up -d)
+(cd infra/migrations && uv sync && uv run alembic upgrade head)
+
+# run everything locally (writes local-only Ed25519 keys into .tmp/)
+bash scripts/dev.sh            # up | logs | stop | status
+cd apps/dashboard && pnpm dev  # console on :3001
+
+# self-hosted Docker path
+bash scripts/selfhost.sh init|up|status|logs|down
+
+# verification (see the ladder in section 7)
+bash scripts/check-consistency.sh          # cross-file docs/code oracle, C1-C10
+cd packages/orchestration && pnpm typecheck && pnpm vitest run
+cd services/<svc> && uv run ruff check . && uv run pytest
+cd apps/dashboard && pnpm typecheck && pnpm test
+
+# operations
+cd packages/orchestration && pnpm scheduler:trigger --list
+scripts/deploy.sh                          # prod (flock + compose pull/up --wait)
+```
+
+---
+
+## 7. Verification ladder
+
+Never claim a change works from a compile or a single unit test. Climb as far as the change warrants:
+
+1. `bash scripts/check-consistency.sh` — the repository's own doc/code oracle (C1–C10).
+2. `uv run ruff check .` in the service you touched (CI enforces this).
+3. `uv run pytest` in every touched service. Note: paper has ~95 test files, evolver ~60; they are the
+   real regression nets for this system.
+4. `pnpm typecheck && pnpm vitest run` in `packages/orchestration` if you touched it; `pnpm typecheck && pnpm test`
+   in the touched app.
+5. **Smoke the real path** — `bash scripts/dev.sh` plus the console or `mastra dev`, then exercise the
+   changed behavior. This repository explicitly separates "code + automated tests" from
+   "real-environment acceptance" (see `docs/04-current-state.md`); leases, fills, and approvals behave
+   differently in a live stack than under fixtures.
+6. Python type checking is **best-effort**, and nothing in this repository claims otherwise: CI runs
+   `uv run mypy . || true` and `CONTRIBUTING.md §9` says "ruff must pass, mypy best-effort". Do not
+   report mypy as passing CI.
+
+---
+
+## 8. Configuration
+
+- **Model**: one root `.env` shared by TypeScript and Python. `scripts/dev.sh` sources it; Python
+  services read it by repo-root path via pydantic-settings; dashboard reads it through
+  `loadRootEnv()`. Legacy per-service `.env` files still act as cwd-level overrides.
+- **Never print, commit, or paste values.** `.env.example`, `infra/.env.*.example` are templates with
+  placeholders only. Keys that matter: `DATABASE_URL`, `JWT_SECRET`, `AUTH_ENABLED`,
+  `LLM_PROVIDER` + provider key, `LLM_CONFIG_ENCRYPTION_KEY`,
+  `EVOLUTION_CREDENTIAL_{PRIVATE,PUBLIC}_KEY_B64` (required for evolution; fail-closed without),
+  `EVENT_EVOLUTION_ENABLED` (E2 flag, default false), `SCHEDULER_ENABLED` (default false),
+  `FRED_API_KEY`, `CONSTITUENT_SNAPSHOT_INDICES`.
+- Environment defaults differ deliberately: `AUTH_ENABLED=false` in the dev template, `true` in the
+  production template. Single-worker assumptions are load-bearing (paper login throttle, live-runner
+  tasks, pending approvals, evolver managers) — only a warning is printed.
+
+---
+
+## 9. Documentation you must not trust
+
+Where documentation and code disagree, the code wins. A read-only audit of HEAD `4822bb9` found the
+misstatements below. Anything marked **fixed** was corrected on branch `docs/agent-onboarding` — read
+that branch's commits for the exact edits, and re-verify any other page against the code.
+
+Still open (deliberately not changed — each needs an owner or design decision):
+
+| Claim | Reality |
+|---|---|
+| Private-ADR policy ("ADRs live in gitignored `docs/miro/`; public docs must not cite them") | `docs/designs/e2-event-evolution-loop.md` is ADR-shaped and `docs/04` cites ADR ids; the consistency check (C2) only scans `docs/00,01,03`, so it cannot catch this |
+| README: "79 factors" | No such count in source; establishing it needs a runtime `GET /catalog` |
+| Release metadata (`CITATION.cff` version, `SECURITY.md` alpha tag) | Tag set stops earlier than claimed — maintainer-owned, do not edit |
+| `packages/orchestration` `pnpm lint` | A stub (`echo 'TODO: add eslint or biome'`); no linter runs in CI for that package |
+| `services/evolver/runtime/executor.py::_service_token` | Signs `{sub, exp}` with no `token_use` / audience / purpose, unlike every other evolver outbound token. Data's `/bars` only needs a valid signature + `sub`, so it works — but such a token is also accepted as a *user* token by any service sharing `JWT_SECRET`. Deviation, not a proven breach; tightening it needs a coordinated data-side change |
+
+Corrected in this branch (listed so nobody re-reports them as open):
+
+| Was | Now |
+|---|---|
+| `docs/01-architecture-overview.md`: agents are orchestrator + trader + risk | single orchestrator agent |
+| `docs/01-architecture-overview.md`: `memory/` is PostgresStore | LibSQL store under `.data/` |
+| Docs/README said 5 lifecycle hook events | six |
+| This onboarding doc claimed `CLAUDE.md` / the PR template make mypy a CI red line | no such claim exists in either file; `CONTRIBUTING.md §9` and CI both say best-effort |
+| `AGENTS.md` pointed at a non-existent `docs/brand/` | now `docs/00-05` |
+| `packages/orchestration/README.md`: orchestrator coordinates trader / risk subagents | single agent |
+| `services/paper/execution/*`: promised a `LiveGateway` | no live gateway; docstring now says so |
+| `services/paper/api/health.py`: "paper does not touch the DB" | paper is DB-backed; `/health` is liveness-only by design |
+| `services/paper/strategy_authoring/fitness.py`: turnover = `turnover / 252` | `(trades / bars_processed) * 100` |
+| `services/paper/kernel/clock.py`: `set_timer` framed as a pending D-5 gap | states it is unimplemented and that the live path drives `TestClock` |
+| `apps/dashboard/src/i18n/routing.ts` pointed at `src/middleware.ts` | `src/proxy.ts` |
+| `services/evolver/.../evaluator/runner.py` implied it used the data service | documents that `bars` is empty; production path is `FrozenDatasetEvaluator` |
+| `services/data/src/inalpha_data/cache_manager.py`: dead module with an import that would raise | deleted |
+
+---
+
+## 10. Risk areas (treat as fragile)
+
+1. `services/evolver` imports `inalpha_paper` in-process (~25 modules) — the only cross-service Python
+   coupling, with no contract test guarding it. Any paper-internal refactor can break evolver at import time.
+2. `services/paper/src/inalpha_paper/live_runner.py` — 1600+ lines of DB + asyncio + risk + plan/exec;
+   a documented session/DB divergence risk and heavy reliance on status guards.
+3. Fail-open HTTP risk path (see ground rule 3).
+4. Single-worker assumptions across four components.
+5. Evolution lease/fencing/budget/holdout semantics — dense, mostly documented in evidence receipts.
+6. Approval identity (`authSub + sessionId + toolName + inputDigest` plus a frozen LLM snapshot):
+   changing input projection either invalidates approved commands or mismatches what was approved.
+7. `MessageBus.publish` to an unmatched topic silently returns 0 — a typo'd topic disables a component
+   with no error.
+8. Hand-written raw-SQL migrations (no autogenerate; revision `0039` missing, `0043` duplicated).
+9. Market→venue routing is prompt prose, not code.
+10. Wallet/accounting: legacy runs are explicitly "unverified or contaminated"; violating the documented
+    lock order risks double-spending capital.
+
+---
+
+## 11. Resolve before assuming
+
+Open items from the familiarization pass that need first-hand confirmation:
+
+- Exact `approval_token` TTL / one-time / quorum mechanics in `paper/storage/trade_plans.py`.
+- Whether the dashboard actually sends the injected LLM-config header and mints grants as the two sides assume.
+- Whether anything sets `strategy_candidates.status = 'rejected'`.
+- Runtime role of `permissions/ask-cache.ts` versus `PendingApprovalsStore`.
+- Whether `UserPromptSubmit` hooks are registered anywhere (declared only, so far).
+- Real runtime health of the stack, and whether the production claims in `docs/04` /
+  `docs/validation/*.json` reproduce.
+
+---
+
+## 12. Definition of done
+
+A change is complete when: the defect or requested change is fixed at its cause (not symptom-suppressed);
+every caller and the corresponding documentation (concept docs, docstrings, generated artifacts) are
+updated in the same change; the verification ladder in section 7 has been climbed as far as the change
+warrants, with the exercised path named; no sensitive material was added anywhere; and the change is
+committed and pushed on a task branch (`docs/*`, `fix/*`, `feature/*`, `chore/*`, `test/*`) with an
+English Conventional Commit message. `main` and `staging` are protected — they take PRs, not direct pushes.
+
+Unrelated defects and observations found along the way: report them, do not silently fix them.
+
+## 13. Further reading
+
+- [`AGENTS.md`](AGENTS.md) — hard constraints, naming, task routing, "do not touch" list
+- [`CLAUDE.md`](CLAUDE.md) — project memory and the local pre-push checklist
+- [`CONTRIBUTING.md`](CONTRIBUTING.md) — branch model, staging rules, commit/PR conventions
+- [`docs/00-context.md`](docs/00-context.md) — goals, boundaries, borrowed designs
+- [`docs/01-architecture-overview.md`](docs/01-architecture-overview.md) — layers and invariants (see section 9 for what is stale)
+- [`docs/03-kernel-design.md`](docs/03-kernel-design.md) — kernel abstractions (Phase C baseline plus deltas)
+- [`docs/04-current-state.md`](docs/04-current-state.md) — authoritative status and priority list
+- [`docs/05-paper-wallet-accounting.md`](docs/05-paper-wallet-accounting.md) — wallet isolation and rollout rules
+- [`docs/usage/README.md`](docs/usage/README.md) — unified LLM usage ledger contract
diff --git a/apps/dashboard/src/i18n/routing.ts b/apps/dashboard/src/i18n/routing.ts
index 1454fb55..19668a87 100644
--- a/apps/dashboard/src/i18n/routing.ts
+++ b/apps/dashboard/src/i18n/routing.ts
@@ -2,7 +2,7 @@ import { defineRouting } from "next-intl/routing";
 
 /**
  * 与官网 apps/web 保持一致:en 默认无前缀,zh 走 /zh。
- * 控制台是动态应用,locale 协商靠 middleware(见 src/middleware.ts)。
+ * 控制台是动态应用,locale 协商靠 middleware(见 src/proxy.ts)。
  */
 export const routing = defineRouting({
   locales: ["en", "zh"],
diff --git a/docs/01-architecture-overview.md b/docs/01-architecture-overview.md
index 38bfa39c..078078e5 100644
--- a/docs/01-architecture-overview.md
+++ b/docs/01-architecture-overview.md
@@ -18,14 +18,14 @@
 ┌───────────────────────────────▼──────────────────────────────────────┐
 │  编排层 · packages/orchestration · Mastra (TypeScript)                 │
 │                                                                        │
-│   agents/      orchestrator → trader / risk（按市场分类自动路由 venue）│
+│   agents/      单个 orchestrator agent（按市场分类自动路由 venue）     │
 │   tools/       data.* web.* factor.* research.* paper.* trade.* evolver.*│
 │                swarm.*                                                  │
 │                + mcp__<server>__<verb>（可插拔外部 MCP）               │
-│   hooks/       5 类生命周期事件 + Stop（PreToolUse / PostToolUse / …） │
+│   hooks/       6 类生命周期事件（PreToolUse / PostToolUse / Stop …）   │
 │   permissions/ allow / ask / deny 三态（deny > allow > ask > default） │
 │   plan/exec    create_plan → approve_plan → execute_plan（一次性 token）│
-│   memory/      PostgresStore · 用户偏好 / 历史会话；plan 由 paper DB 持久化│
+│   memory/      LibSQL 存储（.data/）· 会话与偏好；plan 由 paper DB 持久化  │
 └───────────────────────────────┬──────────────────────────────────────┘
                                 │ HTTP / MCP（每个 tool 调对应服务）
         ┌──────────┬──────────┬──────────┬──────────┬──────────┐
diff --git a/packages/orchestration/README.md b/packages/orchestration/README.md
index 34d10de1..1989fedf 100644
--- a/packages/orchestration/README.md
+++ b/packages/orchestration/README.md
@@ -5,7 +5,8 @@ Mastra / TypeScript 编排层：把 Python services 和外部 MCP 包装成 agen
 
 ## 当前职责
 
-- **Agent 路由**：orchestrator 协调 trader、risk、research-hub，并按市场分类选择数据源。
+- **Agent 路由**：单个 orchestrator agent 直接调用各 tool，并按市场分类选择数据源
+  （trader / risk 子 agent 已在 D-8a′ 移除；下单仍走 plan/exec，模型没有直下单 tool）。
 - **Tool 族**：`data.*` / `web.*` / `factor.*` / `research.*` / `paper.*` / `trade.*` /
   `evolver.*` / `swarm.*` / `skill.*` / `risk.*` / `scheduler.*` / `divination.*` /
   `sandbox.*`，以及可插拔 `mcp__<server>__<verb>`。
diff --git a/packages/orchestration/src/mastra/memory.ts b/packages/orchestration/src/mastra/memory.ts
index 872ef266..36993bd4 100644
--- a/packages/orchestration/src/mastra/memory.ts
+++ b/packages/orchestration/src/mastra/memory.ts
@@ -21,14 +21,14 @@
  * 如果 caller 漏传 ``resourceId``，Mastra 默认会用空字符串 → **不同用户共享同一空
  * 桶** → 历史互窜 + 隐私泄漏。
  *
- * 上 prod 前必须：
+ * 约束：
  *
- * 1. 在 agent.generate / stream 入口（apps/web Next.js API route）强制
- *    ``assertScopedRequest({ resourceId, threadId })``
- * 2. ``resourceId`` 从 session.user.id 派生，**不能**让客户端任意传
- * 3. LibSQLStore 换成 PostgresStore 走与业务表同一 PG（按 user_id 分区）
- *
- * 改成多实例 / 隔离的触发条件：上多用户 / live trading 时按 ADR-0011 §isolation 拆。
+ * 1. 调用方（dashboard BFF → mastra agent 入口）必须传 ``resourceId`` / ``threadId``，
+ *    并自行调用 ``assertScopedRequest``——它不会自动生效
+ * 2. ``resourceId`` 必须来自已认证会话（identity middleware 写入的 sub），**不能**让
+ *    客户端任意传
+ * 3. LibSQL 是单写者文件库：单实例部署够用；要改成多实例 / 按 owner 隔离时再换
+ *    PostgresStore，走与业务表同一 PG
  */
 import { LibSQLStore } from "@mastra/libsql";
 import { Memory } from "@mastra/memory";
diff --git a/services/data/src/inalpha_data/cache_manager.py b/services/data/src/inalpha_data/cache_manager.py
deleted file mode 100644
index d480c166..00000000
--- a/services/data/src/inalpha_data/cache_manager.py
+++ /dev/null
@@ -1,159 +0,0 @@
-"""K 线缓存管理器 —— 增量拉取 + 缓存复用。
-
-功能：
-1. 从 TimescaleDB 读取已缓存的 K 线
-2. 计算缺失的时间范围
-3. 只拉取增量部分
-4. UPSERT 写入缓存（避免重复）
-
-优化效果：
-- 多用户查询同一股票：缓存复用，0 次请求
-- 增量更新：只拉最新数据，节省 80-99% 配额
-
-注意：本模块为基础设施工厂，当前尚未被 backfill/bars API 集成。
-后续 PR 将在 backfill.py 中调用 get_cache_manager() 实现增量缓存。
-"""
-from __future__ import annotations
-
-import threading
-from datetime import datetime
-
-import asyncpg
-from inalpha_shared import get_logger
-
-_logger = get_logger(__name__)
-
-
-class CacheManager:
-    """K 线缓存管理器"""
-
-    def __init__(self, db_pool: asyncpg.Pool):
-        self.pool = db_pool
-
-    async def get_cached_bars(
-        self,
-        venue: str,
-        symbol: str,
-        timeframe: str,
-        from_ts: datetime,
-        to_ts: datetime,
-    ) -> list[dict]:
-        """从缓存读取 K 线"""
-        async with self.pool.acquire() as conn:
-            rows = await conn.fetch(
-                """
-                SELECT ts, open, high, low, close, volume
-                FROM bars
-                WHERE venue = $1
-                  AND symbol = $2
-                  AND timeframe = $3
-                  AND ts >= $4
-                  AND ts <= $5
-                ORDER BY ts ASC
-                """,
-                venue,
-                symbol,
-                timeframe,
-                from_ts,
-                to_ts,
-            )
-
-            return [dict(row) for row in rows]
-
-    async def get_last_cached_ts(
-        self,
-        venue: str,
-        symbol: str,
-        timeframe: str,
-    ) -> datetime | None:
-        """获取最后缓存时间"""
-        async with self.pool.acquire() as conn:
-            row = await conn.fetchrow(
-                """
-                SELECT MAX(ts) as last_ts
-                FROM bars
-                WHERE venue = $1 AND symbol = $2 AND timeframe = $3
-                """,
-                venue,
-                symbol,
-                timeframe,
-            )
-
-            return row["last_ts"] if row and row["last_ts"] else None
-
-    async def upsert_bars(
-        self,
-        venue: str,
-        symbol: str,
-        timeframe: str,
-        bars: list[tuple],
-    ) -> int:
-        """插入新 K 线（UPSERT 避免重复）"""
-        if not bars:
-            return 0
-
-        async with self.pool.acquire() as conn:
-            await conn.executemany(
-                """
-                INSERT INTO bars (venue, symbol, timeframe, ts, open, high, low, close, volume)
-                VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9)
-                ON CONFLICT (venue, symbol, timeframe, ts) DO NOTHING
-                """,
-                [
-                    (venue, symbol, timeframe, b[0], b[1], b[2], b[3], b[4], b[5])
-                    for b in bars
-                ],
-            )
-            # 返回尝试插入的总条数（ON CONFLICT DO NOTHING 可能少于此数）
-            return len(bars)
-
-    async def get_cached_count(
-        self,
-        venue: str,
-        symbol: str,
-        timeframe: str,
-        from_ts: datetime,
-        to_ts: datetime,
-    ) -> int:
-        """获取缓存条数"""
-        async with self.pool.acquire() as conn:
-            row = await conn.fetchrow(
-                """
-                SELECT COUNT(*) as count
-                FROM bars
-                WHERE venue = $1
-                  AND symbol = $2
-                  AND timeframe = $3
-                  AND ts >= $4
-                  AND ts <= $5
-                """,
-                venue,
-                symbol,
-                timeframe,
-                from_ts,
-                to_ts,
-            )
-
-            return row["count"] if row else 0
-
-
-# 全局缓存管理器实例（延迟初始化，线程安全）
-_cache_manager: CacheManager | None = None
-_cache_manager_lock = threading.Lock()
-
-
-def get_cache_manager() -> CacheManager:
-    """获取缓存管理器实例（线程安全）。"""
-    global _cache_manager
-    if _cache_manager is not None:
-        return _cache_manager
-
-    with _cache_manager_lock:
-        # 双重检查：锁内再次检查，防止竞态
-        if _cache_manager is not None:
-            return _cache_manager
-
-        from inalpha_shared.db import get_db_pool
-
-        _cache_manager = CacheManager(get_db_pool())
-        return _cache_manager
\ No newline at end of file
diff --git a/services/evolver/src/inalpha_evolver/evaluator/runner.py b/services/evolver/src/inalpha_evolver/evaluator/runner.py
index 09b08299..5d33a398 100644
--- a/services/evolver/src/inalpha_evolver/evaluator/runner.py
+++ b/services/evolver/src/inalpha_evolver/evaluator/runner.py
@@ -105,12 +105,17 @@ async def evaluate(
         initial_cash: float = 10000.0,
         fee_rate: float = 0.001,
     ) -> EvaluationResult:
-        """评估单个候选策略（E1 需要 data service 运行）。"""
+        """在内存里评估单个候选策略。
+
+        注意：本类尚未接 data service，``bars`` 传空列表，因此**没有真实行情**。
+        E1 生产路径走 ``FrozenDatasetEvaluator``（冻结 bars + manifest 校验），
+        不要用本类做真实评估。
+        """
         symbol = universe[0] if universe else "BTCUSDT"
 
         fn = partial(
             run_engine_in_subprocess,
-            bars=[],  # TODO: 从 data service 拉 bars
+            bars=[],  # 未接 data service；生产路径见 FrozenDatasetEvaluator
             instrument_id=symbol,
             timeframe=timeframe,
             strategy_id=None,
diff --git a/services/paper/src/inalpha_paper/api/health.py b/services/paper/src/inalpha_paper/api/health.py
index ddf9fd0f..02a8a813 100644
--- a/services/paper/src/inalpha_paper/api/health.py
+++ b/services/paper/src/inalpha_paper/api/health.py
@@ -13,7 +13,8 @@
 async def health() -> HealthResponse:
     """liveness。
 
-    paper service D-6 不直接连 DB（只调 data-service），所以不做 DB ping。
-    D-7 持久化 backtest_runs 之后会加 DB ping。
+    paper 的业务数据（accounts / orders / positions / runs / plans / wallets）都在
+    PostgreSQL，但本端点**刻意只做进程探活**：不连 DB、不反映依赖健康，DB 不可用时
+    仍会返回 200。需要依赖健康时另做 readiness 检查。
     """
     return HealthResponse(status="ok", service="paper", version=__version__)
diff --git a/services/paper/src/inalpha_paper/engine/__init__.py b/services/paper/src/inalpha_paper/engine/__init__.py
index 9792bc38..5c7c74a1 100644
--- a/services/paper/src/inalpha_paper/engine/__init__.py
+++ b/services/paper/src/inalpha_paper/engine/__init__.py
@@ -1,7 +1,9 @@
 """引擎：Portfolio + BacktestEngine + Report。
 
-D-5 范围：``BacktestEngine`` 把 D-3 data + D-4 kernel + D-5 execution 串起来跑完整闭环。
-D-6 起：``LiveEngine``（asyncio 事件循环 + LiveGateway）。
+``BacktestEngine`` 把 data + kernel + execution 串起来跑完整回测闭环。
+模拟盘不走独立的 LiveEngine：``LiveRunnerManager`` 用 ``engine/live_session.py`` 把
+已收盘 bar 喂给同一个内核（``TestClock`` 驱动），live runner 的调度、审批与落库在
+``live_runner.py``。
 """
 from .backtest import BacktestEngine
 from .portfolio import Portfolio
diff --git a/services/paper/src/inalpha_paper/execution/__init__.py b/services/paper/src/inalpha_paper/execution/__init__.py
index b2d2393f..200efbb0 100644
--- a/services/paper/src/inalpha_paper/execution/__init__.py
+++ b/services/paper/src/inalpha_paper/execution/__init__.py
@@ -1,15 +1,14 @@
 """执行链：Strategy → RiskEngine → ExecutionEngine → Gateway → Exchange / 真实经纪商。
 
-D-5 范围：
+当前内容：
 
-- ``Gateway`` 抽象 + ``SimulatedExchange``（一并实现 venue 撮合，回测专用）
-- ``RiskEngine`` —— D-5 简化版 pass-through，规则化校验在 ADR-0011 落地（Mastra 编排层）
+- ``Gateway`` 抽象 + ``SimulatedExchange``（一并实现 venue 撮合）
+- ``RiskEngine`` —— 内核内风控守门（Order 进 Execution 前必经）
 - ``ExecutionEngine`` —— Order 状态机管理 + 路由
+- 规则化风控已落地在 ``risk_rules/``：回测路径由 ``RiskEngine`` 执行、HTTP 路径由
+  ``RiskGuard`` 执行并落 ``risk_locks``；运行状态见 ``docs/04-current-state.md``
 
-D-6+ 范围：
-
-- ``LiveGateway`` 接 Binance via CCXT
-- Risk 规则化校验（hook + permission 体系）
+没有 ``LiveGateway``：不接真实经纪商，模拟盘成交走 ``SimulatedExchange`` + DB 侧 executor。
 """
 from .exchange import EXECUTION_ENGINE_ENDPOINT, SimulatedExchange
 from .execution_engine import ExecutionEngine
diff --git a/services/paper/src/inalpha_paper/execution/gateway.py b/services/paper/src/inalpha_paper/execution/gateway.py
index c9af70cc..e8dd6a3a 100644
--- a/services/paper/src/inalpha_paper/execution/gateway.py
+++ b/services/paper/src/inalpha_paper/execution/gateway.py
@@ -1,13 +1,13 @@
 """Gateway 抽象 —— 经纪商 / 交易所接入。
 
-设计依据 [refs/vnpy.md §3 §7](../../../../docs/refs/vnpy.md)，把 vnpy 的 7 个抽象方法
-裁剪到 MVP 需要的：
+设计依据 vnpy 的 Gateway 抽象（§3 §7），把它的 7 个抽象方法裁剪到当前需要的：
 
 - ``send_order`` —— 唯一必须实现
 - ``cancel_order`` / ``modify_order`` —— 可选（不实现就 raise NotImplementedError）
 
-D-5 起只有 ``SimulatedExchange`` 一个实现（同时承担 venue 撮合职责）。D-6 起加
-``LiveGateway``（CCXT Binance）。
+当前只有 ``SimulatedExchange`` 一个实现（同时承担 venue 撮合职责）。没有连接真实经纪商
+的 gateway：模拟盘成交由 ``SimulatedExchange`` + DB 侧 ``OrderExecutor`` 完成，回测与
+模拟盘共用同一份策略代码，只换 Clock。
 """
 from __future__ import annotations
 
diff --git a/services/paper/src/inalpha_paper/kernel/clock.py b/services/paper/src/inalpha_paper/kernel/clock.py
index 9b0996c3..9aa7921b 100644
--- a/services/paper/src/inalpha_paper/kernel/clock.py
+++ b/services/paper/src/inalpha_paper/kernel/clock.py
@@ -146,11 +146,12 @@ def now_ns(self) -> int:
 
     def set_timer(self, name: str, interval_ns: int, callback: TimerCallback) -> None:
         raise NotImplementedError(
-            "LiveClock.set_timer requires asyncio loop, attached in Engine (D-5)"
+            "LiveClock.set_timer is not implemented; the live path drives TestClock "
+            "from closed bars instead (see engine/live_session.py)"
         )
 
     def cancel_timer(self, name: str) -> None:
-        # D-5 起会实现，目前 noop（保持接口对称）
+        # 未实现定时器，保持接口对称；实时路径不经 timer
         pass
 
 
diff --git a/services/paper/src/inalpha_paper/strategy_authoring/fitness.py b/services/paper/src/inalpha_paper/strategy_authoring/fitness.py
index 1f5baba0..01d20c51 100644
--- a/services/paper/src/inalpha_paper/strategy_authoring/fitness.py
+++ b/services/paper/src/inalpha_paper/strategy_authoring/fitness.py
@@ -7,7 +7,7 @@
 
     fitness = sharpe
             + 0.3 * calmar
-            - 0.10 * turnover_penalty       # turnover / 252（年化换手）
+            - 0.10 * turnover_penalty       # 每根 bar 平均交易数 × 100
             - 1.0 * (drawdown > 0.30)       # 30% 回撤一票否决
 
 MVP 留空：``capacity_penalty``（需 ADV 数据，D-10+ 引入）。