[Bug]: Fenced goal text hides generated Todo sections

[ehz0ah]

Preflight

• Searched existing issues, discussions, branches, and PRs.
• This report concerns reproducible parsing behavior, not an unpatched security vulnerability.
• Contains no credentials, private state, or local paths.

Issue origin

Observed and reproduced in a real environment, then reduced to synthetic input.

Version and area

Commit 741d2e3c0162f785eb21a5fe609f9d93e0eb470b; control plane, goal bootstrap and Todo readback. Not host-specific. Reproduced with Python 3.13.12 on macOS.

Problem

A goal prompt enclosed in a Markdown code fence can produce state whose generated Todos are invisible to the shared parser. Start-goal collapses whitespace, so an otherwise closed multiline fence becomes one line. render_state_markdown inserts that line directly under ## Objective. The parser treats it as an opening fence and hides the remaining document.

Minimal reproduction

Run from a source checkout with Python 3.11 or later. This uses temporary synthetic state only.

from pathlib import Path
from tempfile import TemporaryDirectory
from loopx.bootstrap import render_state_markdown
from loopx.control_plane.todos.active_state_todo_parser import parse_todo_source

with TemporaryDirectory() as directory:
    state = render_state_markdown(
        project=Path(directory),
        goal_id="objective-example",
        adapter_kind="read_only_project_map_v0",
        objective="```text Implement and validate the task. ```",
        updated_at="2026-09-15T00:00:00+00:00",
        goal_doc=None,
        execution_profile=None,
    )
    items, archive, sources = parse_todo_source(state)
    print(sources)
    print(len(items["agent"]))

Expected behavior

Both generated Todo source sections remain readable. The generated connection-validation Todo is present. Objective examples are text, not additional authoritative Todo sections.

Actual behavior

The reproduction prints {'user': None, 'agent': None} and 0.

The focused regression also fails for a tilde fence and an open HTML comment. An unfenced Todo example in the objective is read as an additional Todo. Plain text and a properly closed multiline fenced example pass.

Proposed scope and validation

Isolate objective content in generated Markdown, preserve its text, and verify the direct objective readback path. Inspect both loopx/bootstrap.py:render_state_markdown and loopx/control_plane/projects/registry.py:_state_markdown. Do not weaken fence handling, change Todo authority, or migrate live goal state as a test.

Current regression result: four failing boundary cases, two passing controls, and one passing existing bootstrap test. Follow-up validation should cover bootstrap, project registration, Todo readback, and objective context readback.

Related merged PRs #4299 and #4097 address document ownership and canonical Todo recovery. Neither changes these objective writers.

[Exusty]

认领这个切片（claiming this slice），PR: #4435

方案 / Approach：按 issue 里的建议方向——把生成的 objective 文本隔离起来，而不是弱化 fence 处理。新增 loopx:objective-v0 begin/end 标记，render_state_markdown 与 _state_markdown 都用它包住 objective，visible_markdown_lines 把区内行当散文处理（区外真围栏照旧解析）。同时加 read_objective_text，让 dashboard 的 objective 回读能原样取回以 <!-- 开头的文本。

Scope：两个 writer + 一个 reader + 新测试，不改 Todo 权限语义，不迁移真实 goal state。老 state 文档无标记时行为完全不变。

Note on IP/importance：不适用（本 PR 是 bug fix）。
Validation：新增 tests/control_plane/test_objective_todo_visibility.py（39 例，覆盖 fence / tilde fence / 未闭合 HTML 注释 / 类 Todo 文本 + 纯文本与闭合围栏对照组，两个 writer 各跑一遍）。相关回归 344 过 4 挂，4 个均为本地 [sqlite] 运行时准入失败（Node 22.22.2 < 要求的 22.22.3），已在 PR 中标注为环境阻塞。