Pin initial context readers to resolved selections

docs/reference/required-work-context.md

@@ -196,14 +196,21 @@ marked `required=false` defer their guidance/templates; their routes stay visibl
 and the complete captured attachment must be read before use. Unknown plan
 versions remain whole. Initial pages do not repeat pointers already present in
 their content; the selection digest still prevents a full-input receipt.
-Its `context-read --initial-read` uses the same exact-address selector; paged
-continuations retain the short option and pin the selection digest. A changed
-profile rejects rather than mixing views. Initial navigation can replace fully
+Its `context-read --initial-read` uses the same exact-address selector. New private
+selection references pin the resolved addresses and complete selection digest;
+continuations use the existing explicit reader instead of rerunning initial
+navigation. A later navigation change cannot replace the saved selection.
+Historical recipe-only references and explicit-file initial cursors still reject
+a changed selection digest; recover through the same original full-capture
+reference, without another admission or rewriting the saved descriptor. These
+legacy readers still require reconciliation of their unfinished original Turns
+before retirement; this slice retains their decoder and full-capture recovery.
+Initial navigation can replace fully
 selected children with their complete captured parent to avoid repeating addresses.
 An unselected child, including optional or unknown material, prevents that merge;
 the selected facts and obligations stay identical. Explicit selectors and index
-entries keep their original addresses. The option is read-only and exclusive with explicit
-selectors or discovery; unknown schemas use generic discovery.
+entries keep their original addresses. The option is read-only and exclusive with
+explicit selectors or discovery; unknown schemas use generic discovery.
 The concrete first read replaces the broad scalar preview and root index when it
 fits; unknown schemas or a tight reference budget keep generic discovery. Warm
 deltas keep their existing reader, and disabled native delivery is unchanged.
@@ -232,7 +239,11 @@ activation defaults, saved files, Delta and full-read receipts are unchanged.
 后续偏好相关外部动作前精确作用域新读；偏好及其交付均不授予权限。
 未知 schema、过小引用预算和 warm delta 保留既有路线，
 关闭原生传输的输出不变；完整恢复和原 SHA 校验均保留。只读的 `--initial-read` 不与
-显式选读或索引混用；续页保留短选项和固定的选读摘要，profile 改变则拒绝混读。
+显式选读或索引混用。新私有选读引用固定已解析路径和完整选读摘要，续页复用显式
+reader，不重新运行首读导航；后续导航改变不能替换原选读。历史仅保存导航配方的引用
+及显式文件首读游标在选读摘要变化时仍拒绝，沿同一完整原 capture 引用恢复，不重新
+准入或改写旧描述文件。退役这些 legacy reader 前仍须对账原未完成 Turn；本切片
+保留历史解码与完整捕获恢复。
 首读可把已经完整选中的同级内容合并为其完整父节点，减少地址重复；任何未选中的
 可选或未知内容都会阻止合并，选中事实和义务不变。显式选读和索引仍保留原地址。
 首读也给出现有验收材料地址和当前结算义务。仅明确 `required=false` 的附件延后完整

loopx/cli_commands/quota_capture.py

@@ -292,8 +292,11 @@ def read_context_capture(args: argparse.Namespace, *, presentation: dict[str, An
             selectors = [item["pointer"] for item in views]
             if reference_file is not None and pinned_selection is None:
                 try:
+                    # Pin the resolved addresses, not the initial navigation
+                    # recipe. Existing explicit readers can replay these exact
+                    # bytes after a navigation change, with both digests checked.
                     reference_file = write_context_read_reference(path, digest, selection={
-                        "select": None if initial_read else selectors, "initial_read": initial_read,
+                        "select": selectors, "initial_read": False,
                         "list_fields": list_fields, "list_depth": list_depth, "sha256": expected})
                 except OSError:
                     # A verified read must not require a writable capture root.

tests/control_plane/test_context_read_reference.py

@@ -132,6 +132,58 @@ def changed(operation, **kwargs):
     assert "content" not in rejected
 
 
+@pytest.mark.parametrize("legacy", [False, True], ids=["pinned-addresses", "historical-profile"])
+def test_initial_reference_navigation_change_preserves_original_read_boundary(tmp_path, monkeypatch, legacy):
+    payload, machine, model = fixture(tmp_path)
+    expected = read(model["first_read_command"])
+    assert expected["ok"] and expected["eof"]
+    first = read(model["first_read_command"], max_page_output_bytes=3000, page_bytes=3000)
+    assert first["ok"] and not first["eof"]
+    reference = Path(first["reference_file"])
+    if legacy:
+        # Historical documents stored the navigation recipe, not its addresses.
+        _, descriptor = read_context_read_reference(str(reference))
+        reference = write_context_read_reference(Path(machine["path"]), machine["sha256"], selection={
+            **descriptor["selection"], "select": None, "initial_read": True})
+    reference_before, capture_before = reference.read_bytes(), Path(machine["path"]).read_bytes()
+    from loopx.control_plane.quota import turn_envelope
+
+    original = turn_envelope._turn_envelope_result
+
+    def changed_navigation(operation, **kwargs):
+        if kwargs.get("initial_read"):
+            kwargs = {**kwargs, "initial_read": False, "select": ["/criterion"]}
+        return original(operation, **kwargs)
+
+    monkeypatch.setattr(turn_envelope, "_turn_envelope_result", changed_navigation)
+    # A fresh initial read now has different navigation over the same source.
+    current = read(model["first_read_command"])
+    assert json.loads(current["content"])["selections"] == [{"pointer": "/criterion", "value": payload["criterion"]}]
+    following = read(first["next_command"], reference_file=str(reference))
+    if legacy:
+        assert following["ok"] is False and "digest changed" in following["reason"]
+        assert "content" not in following
+        recovered = read(model["recovery"]["command"], select=["/goal_boundary", "/selected_todo"])
+        assert recovered["ok"] and recovered["eof"]
+        values = {row["pointer"]: row["value"] for row in json.loads(recovered["content"])["selections"]}
+    else:
+        parts = [first["content"]]
+        while True:
+            assert following["ok"] and following["schema_version"] == "context_capture_selection_page_v1"
+            assert _context_pages(following) == []
+            assert following["offset"] == len("".join(parts).encode())
+            parts.append(following["content"])
+            if following["eof"]:
+                break
+            following = read(following["next_command"])
+        assert "".join(parts) == expected["content"]
+        values = {row["pointer"]: row["value"] for row in json.loads("".join(parts))["selections"]}
+    assert values["/goal_boundary"] == payload["goal_boundary"]
+    assert values["/selected_todo"] == payload["selected_todo"]
+    assert reference.read_bytes() == reference_before
+    assert Path(machine["path"]).read_bytes() == capture_before
+
+
 @pytest.mark.parametrize("selection", ["fields", "index", "initial"])
 def test_selected_read_survives_unwritable_reference_metadata(tmp_path, monkeypatch, selection):
     import errno