fix(runtime): defer unselected context and retirement implementations

docs/architecture/rfcs/monorepo-distribution-split-v0.md

@@ -3677,3 +3677,63 @@ closure, paired complete costs, progressive product control, R1/M3 and original
 Goal regression remain unqualified. Keep the nine original requirements,
 previous failed attempts and frozen budget failures; no minimal preset,
 provider default, package publication or paid operation is promoted here.
+
+### 2026-11-13 — Full Turn defers unselected provider implementations
+
+The installed core Turn previously loaded BM25 and OpenViking while resolving
+Reward Memory configuration, and loaded the anti-entropy inspector even when
+retirement governance was disabled. Context-provider exports and the existing
+builtin builder registry now resolve implementations on explicit use. A pure
+OpenViking scope owner is shared by binding normalization and the provider;
+the original URI grammar, public aliases and Python serialization identity are
+retained. Anti-entropy still uses the existing typed activation and applicable
+basis; its feedback builder resolves only after those checks. No capability,
+provider default, authority, claim, stop or settlement rule is added or relaxed.
+
+The 12 fresh-process checks cover registry rejections/custom builders, real
+local BM25 retrieval, binding classification, disabled/enabled retirement and
+attribute/from/star exports. The original 11-check corpus fails five loading
+oracles on immutable base `441c6077e7c5c8e896ea79562761fd483baf157e`, while its
+six export checks pass. Sixteen URI success/rejection cases and protocol-4
+pickle bytes match between installed base and candidate; cross-version reads
+pass both ways. Definitions move to the pure scope source file: declaration
+source introspection is not preserved by the original-module alias.
+
+The final independent Python 3.11 core wheel contains 1886 runtime files that
+match source bytes (SHA256
+`bc36c92dd437b9418e21031e185fefd04984c91a30169cbd866a3c712d92b4fa`).
+Source capability regression passes 628 cases. Real isolated File and SQLite
+execution commits a checked artifact and one settlement; original-key replay
+has no Host, state or quota effects. Instrumented full-Turn module counts,
+including the package root, change from 525 to 516 on execution and 438 to 432
+on replay; capability counts change from 111 to 102 and 56 to 50. Execution
+RPC counts remain 221/220 and replay remains 11. The unselected OpenViking,
+BM25 and anti-entropy inspection implementations disappear from those paths.
+These counts describe the observed Python path, not the complete minimal closure.
+
+Matched installed observations use macOS 27 arm64, Python 3.11.15, Node 26.0.0
+and five
+alternating base/candidate pairs per backend and phase. Each invocation starts
+a new Python process. Before first execution, the isolated managed RPC server
+is normally stopped and must restart; completed-Turn replay reuses that server.
+Setup and independent readback are outside timing; OS, bytecode and compilation
+caches are not reset. Nearest-rank p95 is the maximum of five samples.
+
+| Backend / phase | Wall median / p95 ms, base → candidate | Median invocation RSS MiB, base → candidate |
+| --- | --- | --- |
+| SQLite / fresh execution | 2922.5 / 3561.9 → 3006.8 / 4139.9 | 81.81 → 81.25 |
+| SQLite / completed-Turn replay | 809.6 / 948.6 → 889.0 / 1415.0 | 44.73 → 44.42 |
+| File / fresh execution | 2858.7 / 3005.5 → 2862.5 / 2985.0 | 81.09 → 81.36 |
+| File / completed-Turn replay | 645.4 / 679.7 → 692.5 / 773.9 | 44.73 → 44.28 |
+
+All 40 invocations preserve the same independent result/effect fingerprint;
+stdout sizes remain 4002 bytes for execution and 4548 for replay. The candidate
+has slower SQLite medians and a higher tail: no latency or complete-cost
+improvement is qualified. Invocation RSS excludes the persistent Node server;
+zero block-I/O counters do not prove zero file access. Separate controlled
+observers verify Python import/open/process/loopback events and Node module/
+loopback/exit events, but do not provide OS-wide closure, complete IPC/I/O or
+model-token, service-cost and human-attention accounting. The Host remains
+synthetic. Full V1–V5/V10, real model adoption/recovery, progressive product
+control, R1/M3, independent Standard packages and original-goal regression remain
+open. Preserve all nine original requirements and earlier frozen failures.

docs/architecture/rfcs/monorepo-distribution-split-v0.zh-CN.md

@@ -2866,3 +2866,52 @@ journal 使用的 transaction-key 预约保留精确恢复兼容分支；已启
 实际模型采用、完整 V1–V5/V10、默认闭包、配对完整成本、渐进产品控制、R1/M3 与
 原始 Goal 回归仍未验收。保留原九项要求、此前失败尝试和冻结预算失败；本次没有
 推广 minimal preset、provider 默认、包发布或付费操作。
+
+### 2026-11-13 — 完整 Turn 延后加载未选择的 Provider 实现
+
+此前独立 core 的已安装 Turn 在解析 Reward Memory 配置时加载 BM25 和
+OpenViking，并在退休治理关闭时加载 anti-entropy 检查器。现在 context-provider
+公开导出和原 builtin builder registry 在明确使用时解析实现。绑定归一化与实际
+provider 共用纯 OpenViking scope owner，原 URI 语法、公开别名与 Python 序列化
+身份保留。Anti-entropy 仍使用原 typed 激活规则和适用 basis，通过这些检查后才
+解析 feedback builder。没有新增或放宽 capability、provider 默认值、authority、
+claim、stop 或 settlement 规则。
+
+12 项新进程检查覆盖 registry 拒绝／custom builder、真实本地 BM25 检索、绑定
+分类、关闭／开启退休治理，以及 attribute/from/star 导出。原 11 项检查在不可变
+基线 `441c6077e7c5c8e896ea79562761fd483baf157e` 上有五项加载 oracle 失败，
+六项公开导出检查通过。两个安装版本的 16 类 URI 成功／拒绝结果与 protocol-4
+pickle 字节一致，跨版本双向读取通过。定义移到纯 scope 源文件，旧模块别名
+不能保留声明源码的 introspection 行为。
+
+最终独立 Python 3.11 core wheel 的 1886 个 runtime 文件与源码字节一致，SHA256
+为 `bc36c92dd437b9418e21031e185fefd04984c91a30169cbd866a3c712d92b4fa`。
+源码 capability 回归通过 628 项。真实隔离 File 与 SQLite 执行提交已检查产物，
+且只结算一次；原 key 重放没有 Host、state 或 quota 副作用。含 package root 的
+受控完整 Turn 模块数在执行时由 525 降到 516，重放由 438 降到 432；capability
+模块数分别由 111 降到 102、56 降到 50。执行 RPC 数仍为 221/220，重放仍为 11。
+这些路径中不再出现未选择的 OpenViking、BM25 和 anti-entropy 检查实现。计数只
+描述已观测的 Python 路径，不能代表完整最小闭包。
+
+配对安装测量使用 macOS 27 arm64、Python 3.11.15、Node 26.0.0，每个 backend、
+阶段各交替运行
+五对基线／候选。每次调用启动新 Python 进程；首次执行前通过原 owner 正常停止
+隔离的 managed RPC server，使其重新启动，已完成 Turn 重放则复用该 server。
+准备与独立回读不计入耗时，OS、bytecode 和 compilation cache 不重置。
+Nearest-rank p95 等于五个样本的最大值。
+
+| Backend／阶段 | Wall 中位数／p95 ms，基线 → 候选 | 调用 RSS 中位数 MiB，基线 → 候选 |
+| --- | --- | --- |
+| SQLite／首次执行 | 2922.5 / 3561.9 → 3006.8 / 4139.9 | 81.81 → 81.25 |
+| SQLite／已完成 Turn 重放 | 809.6 / 948.6 → 889.0 / 1415.0 | 44.73 → 44.42 |
+| File／首次执行 | 2858.7 / 3005.5 → 2862.5 / 2985.0 | 81.09 → 81.36 |
+| File／已完成 Turn 重放 | 645.4 / 679.7 → 692.5 / 773.9 | 44.73 → 44.28 |
+
+40 次调用的独立结果／副作用指纹一致，stdout 大小保持首次执行 4002 bytes、
+重放 4548 bytes。候选 SQLite 中位数较慢、尾延迟较高，本轮不取得延迟或完整成本
+改善资格。调用 RSS 不包含持久 Node server，block-I/O 计数为零也不能证明没有
+文件访问。独立受控 observer 验证 Python import/open/process/loopback 事件，以及
+Node module/loopback/exit 事件，但不能提供 OS 级完整闭包、完整 IPC/I/O、model
+token、service cost 或人工注意力核算。Host 仍为 synthetic。完整 V1–V5/V10、
+真实模型采用／恢复、渐进产品控制、R1/M3、独立 Standard packages 与原目标回归
+仍开放，保留原九项要求和此前冻结失败。

loopx/capabilities/anti_entropy/__init__.py

@@ -1,30 +1,43 @@
 from __future__ import annotations
 
-from .agent_tool import anti_entropy_inspect
-from .catalog_entry import ANTI_ENTROPY_CATALOG_ENTRY
-from .cli import run_anti_entropy_verify
-from .contract import (
-    ANTI_ENTROPY_CAPABILITY_ID,
-    ANTI_ENTROPY_RECEIPT_SCHEMA_VERSION,
-    ANTI_ENTROPY_RETIREMENT_BASIS_SCHEMA_VERSION,
-    AntiEntropyFinding,
-    AntiEntropyReceipt,
-    RetirementBasis,
-)
-from .feedback import (
-    ANTI_ENTROPY_INSPECT_TOOL_DESCRIPTOR,
-    build_anti_entropy_turn_context,
-    run_anti_entropy_feedback_cycle,
-)
-from .oracles import (
-    audit_ghost_code,
-    evaluate_diff_hygiene,
-    verify_delete_first,
-)
-from .turn_context import (
-    anti_entropy_context_registration,
-    anti_entropy_context_result,
-)
+from importlib import import_module
+from typing import Any
+
+# Importing the Turn registration or contract is not a request to run the
+# inspector. Explicit compatibility exports still resolve to their owners.
+_EXPORTS = {
+    "anti_entropy_inspect": "agent_tool",
+    "ANTI_ENTROPY_CATALOG_ENTRY": "catalog_entry",
+    "run_anti_entropy_verify": "cli",
+    "ANTI_ENTROPY_CAPABILITY_ID": "contract",
+    "ANTI_ENTROPY_RECEIPT_SCHEMA_VERSION": "contract",
+    "ANTI_ENTROPY_RETIREMENT_BASIS_SCHEMA_VERSION": "contract",
+    "AntiEntropyFinding": "contract",
+    "AntiEntropyReceipt": "contract",
+    "RetirementBasis": "contract",
+    "ANTI_ENTROPY_INSPECT_TOOL_DESCRIPTOR": "feedback",
+    "build_anti_entropy_turn_context": "feedback",
+    "run_anti_entropy_feedback_cycle": "feedback",
+    "audit_ghost_code": "oracles",
+    "evaluate_diff_hygiene": "oracles",
+    "verify_delete_first": "oracles",
+    "anti_entropy_context_registration": "turn_context",
+    "anti_entropy_context_result": "turn_context",
+}
+
+
+def __getattr__(name: str) -> Any:
+    module = _EXPORTS.get(name)
+    if module is None:
+        raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+    value = getattr(import_module(f".{module}", __name__), name)
+    globals()[name] = value
+    return value
+
+
+def __dir__() -> list[str]:
+    return sorted(set(globals()) | set(__all__))
+
 
 __all__ = [
     "ANTI_ENTROPY_CAPABILITY_ID",

loopx/capabilities/anti_entropy/turn_context.py

@@ -3,13 +3,14 @@
 from __future__ import annotations
 
 import json
+import sys
 from collections.abc import Mapping
+from importlib import import_module
 from pathlib import Path
 from typing import Any
 
 from ...control_plane.capability_hooks import TurnContextHookRegistration
 from .contract import ANTI_ENTROPY_CAPABILITY_ID, RetirementBasis
-from .feedback import build_anti_entropy_turn_context
 
 ANTI_ENTROPY_CONTEXT_MAX_RESULT_BYTES = 64 * 1024
 _BASIS_KEYS = ("retirement_basis", "anti_entropy_basis")
@@ -26,6 +27,18 @@
 _UNSET = object()
 
 
+def __getattr__(name: str) -> Any:
+    if name != "build_anti_entropy_turn_context":
+        raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+    value = getattr(import_module(".feedback", __package__), name)
+    globals()[name] = value
+    return value
+
+
+def __dir__() -> list[str]:
+    return sorted(set(globals()) | {"build_anti_entropy_turn_context"})
+
+
 def _embedded_json_objects(text: str):
     """Yield JSON objects embedded in free text, outermost first.
 
@@ -484,7 +497,7 @@ def anti_entropy_context_result(
     )
     is_recovery = bool(recovery_evidence)
 
-    context = build_anti_entropy_turn_context(
+    context = getattr(sys.modules[__name__], "build_anti_entropy_turn_context")(
         basis=basis,
         goal_id=goal_id,
         workspace=workspace,

loopx/capabilities/context_providers/__init__.py

@@ -1,27 +1,40 @@
-"""Reusable bounded context-provider integrations for LoopX capabilities."""
-
-from .base import (
-    canonical_context_text,
-    canonical_context_lines,
-    canonical_context_matches,
-    ContextProviderItem,
-    ContextProviderRetrieval,
-    ContextProviderSync,
-)
-from .factory import build_context_provider, register_context_provider
-from .bm25 import BM25ContextProvider, build_bm25_context_provider
-from .openviking import (
-    OpenVikingContextProvider,
-    build_openviking_context_provider,
-)
-from .service_ownership import (
-    context_provider_service_restarted,
-    load_context_provider_service_ownership,
-)
-
-
-register_context_provider("openviking", build_openviking_context_provider)
-register_context_provider("bm25", build_bm25_context_provider)
+"""Context-provider exports resolve without selecting unrelated implementations."""
+
+from __future__ import annotations
+
+from importlib import import_module
+from typing import Any
+
+_EXPORTS = {
+    "canonical_context_text": "base",
+    "canonical_context_lines": "base",
+    "canonical_context_matches": "base",
+    "ContextProviderItem": "base",
+    "ContextProviderRetrieval": "base",
+    "ContextProviderSync": "base",
+    "build_context_provider": "factory",
+    "register_context_provider": "factory",
+    "BM25ContextProvider": "bm25",
+    "build_bm25_context_provider": "bm25",
+    "OpenVikingContextProvider": "openviking",
+    "build_openviking_context_provider": "openviking",
+    "context_provider_service_restarted": "service_ownership",
+    "load_context_provider_service_ownership": "service_ownership",
+}
+
+
+def __getattr__(name: str) -> Any:
+    module = _EXPORTS.get(name)
+    if module is None:
+        raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+    value = getattr(import_module(f".{module}", __name__), name)
+    globals()[name] = value
+    return value
+
+
+def __dir__() -> list[str]:
+    return sorted(set(globals()) | set(_EXPORTS))
+
 
 __all__ = [
     "ContextProviderItem",

loopx/capabilities/context_providers/factory.py

@@ -7,7 +7,26 @@
 
 
 ContextProviderBuilder = Callable[[Mapping[str, Any]], ContextProvider]
-_CONTEXT_PROVIDER_BUILDERS: dict[str, ContextProviderBuilder] = {}
+
+
+def _build_openviking(config: Mapping[str, Any]) -> ContextProvider:
+    from .openviking import build_openviking_context_provider
+
+    return build_openviking_context_provider(config)
+
+
+def _build_bm25(config: Mapping[str, Any]) -> ContextProvider:
+    from .bm25 import build_bm25_context_provider
+
+    return build_bm25_context_provider(config)
+
+
+# Builtin ids are already registered, so extensions retain the same duplicate
+# rejection. Resolving that registry does not select either provider's code.
+_CONTEXT_PROVIDER_BUILDERS: dict[str, ContextProviderBuilder] = {
+    "openviking": _build_openviking,
+    "bm25": _build_bm25,
+}
 
 
 def register_context_provider(