[gcl-coder]

What this changes

session_dash._status_bucket folds a goal's free-form status and waiting_on
into one of five operator buckets (active / needs_user / blocked / done
/ other). It decided that fold by testing marker words as substrings
(marker in status_text), and _USER_WAITING_MARKERS did not list
controller even though the runtime emits it as a human gate.

Two consequences, both reproduced below:

1. A goal waiting on the controller is not counted as needs-you.
2. A status that merely contains a marker's letters lands in the wrong
   bucket: installed reads as stalled, incomplete reads as complete,
   undone reads as done.

This PR matches markers as whole words, treats a negated status
(not_completed) as not-done, and adds controller to the human-gate
vocabulary. It adds the first test file for this module, which had none.

Why (evidence)

loopx/control_plane/runtime/session_runtime.py:76 builds an attention item
with status="unregistered_runtime_goal", waiting_on="controller".
loopx/control_plane/goals/acceptance_observation.py:186 and
loopx/control_plane/handoff/project_agent_context.py:240 both read
{"controller", "user_or_controller"} as the human-gate set. session_dash
was the only reader that disagreed, so the same gate showed up as
needs_user or other depending on which emitter wrote it.

Measured on this tree before the change:

_status_bucket(status="unregistered_runtime_goal", waiting_on="controller")     -> other        # should be needs_user
_status_bucket(status="unregistered_runtime_goal", waiting_on="user_or_controller") -> needs_user
_status_bucket(status="installed",   waiting_on=None) -> blocked   # 'stalled' in 'installed'
_status_bucket(status="uninstalled", waiting_on=None) -> blocked
_status_bucket(status="incomplete",  waiting_on=None) -> done      # 'complete' in 'incomplete'
_status_bucket(status="undone",      waiting_on=None) -> done
_status_bucket(status="not_completed", waiting_on=None) -> done

After the change every one of those is other, and the buckets that were
already right are unchanged (done / completed / closed / shipped /
archived / blocked_by_safety / stalled / failed /
active-progress / in_progress / running / keepalive / relaunched).

The projection is rendered by loopx dash and dash_server, so the overview
strip's needs-you count and each goal's status badge were reading it.

What did not change

• _todo_tally still compares a todo status by exact membership in
  _DONE_STATUS_MARKERS; that path was already exact and its markers are
  untouched.
• Bucket precedence is unchanged: done -> blocked -> needs_user ->
  active -> other.
• No new input is read, no field is added to the projection, and
  session_dash_projection_v1 keeps its shape.

Tests

New tests/presentation/test_session_dash_status_bucket.py (the module had no
tests at all before this PR):

• the documented and in-tree statuses keep their bucket,
• a marker inside a longer word does not set a bucket,
• controller and user_or_controller both land in needs_user, while
  agent / codex / auto / external_evidence do not,
• a blocked status still outranks the waiting reason,
• end to end through build_session_dash_projection: a controller-gated goal
  is counted under overview.goals_by_status.needs_user, and an installed
  goal is not counted as blocked.

Against the code on main the new file fails 4 of its 16 cases; against this
branch all 16 pass.

Scope note

This is the same read-a-word-as-a-substring defect class reported in #4336 one
layer down (a prose matcher reading installed as stalled). #4336's fix went
into the quota path; this closes the presentation reader that still had it. It
does not change, and is not required for, that issue's typed-observation
contract.

[gcl-coder]

CI note: the windows-powershell failure is a 100 s deadline, not a wrong value

windows-powershell fails at step 16, Qualify original-task recovery and interactive input on Windows:

tests/test_delegation_task_revalidation.py::test_revalidation_preserves_an_internal_lock_timeout[file]
FAILED  ... - Failed: {'operation_id': 'analysis-1', 'request_id': '9328474d…', 'agent_id': 'analyst', 'todo_id': …

1 failed, 13 passed, 5 skipped, 6 deselected in 183.91s

The line pytest echoes is assert wait(runner)["status"] == "rejected", but the failure message is
not an AssertionError — it is the dict that wait() prints when it gives up. wait() is
tests/test_local_delegation.py:262-272:

def wait(service, operation="analysis-1"):
    deadline = time.monotonic() + 100
    while time.monotonic() < deadline:
        result = service.read(operation)
        if result["status"] in {"accepted", "rejected"}:
            if not result["worker_active"]:
                return result
        elif result.get("error"):
            pytest.fail(str(result))
        time.sleep(0.25)
    pytest.fail(str(service.read(operation)))

So the delegation run simply never reached a terminal state within its own 100 s budget on this
runner. No assertion compared a wrong value; the 14 cases in that step consumed 183.91 s.

Why this cannot come from this branch

This PR changes exactly two files:

• loopx/presentation/projections/session_dash.py
• tests/presentation/test_session_dash_status_bucket.py (new, 16 cases)

tests/test_delegation_task_revalidation.py drives a detached delegation worker over mcp stdio;
it does not import the presentation projection, and no path from the delegation runner reaches
session_dash. The test itself is upstream code (tests: qualify original task revalidation,
4184c2b33 fix(collaboration): recover original tasks without worker or stdin races (#6866)).

What is green on this head

build, Sign-off, kernel-contracts, kernel-static-checks, chat-bundle, changes,
dependency-review, node-minimum-compatibility, dashboard-acceptance, typescript-coverage,
stage2c-correctness-e2e, stage2c (mutants 0 / installed 0 / e2e 1 / e2e 2), all three
chat-bundle-browser shards, all three dashboard-browser shards, all three typescript-core
shards. The test-shard matrix never left pending on this run, so those shards have not reported
either way yet.

Local verification on this branch

Python 3.14.0, pytest 9.2.0, ruff 0.17.3:

$ python -m pytest -q tests/presentation/test_session_dash_status_bucket.py
16 passed in 0.02s

$ python -m pytest -q tests/presentation
50 passed in 2.17s

$ ruff check loopx/presentation/projections/session_dash.py tests/presentation/test_session_dash_status_bucket.py
All checks passed!

$ python examples/session-dash-panel-smoke.py
session dash panel smoke: OK

Against main the same 16 cases fail 4 of 16, which is what makes them regression tests rather
than characterization tests.

[gcl-coder]

Follow-up: what the rerun showed

Reopening this PR (same head 64975e2f9, no code change) turned the windows-powershell red into a pass. That confirms the earlier reading: it was the 100 s deadline inside wait(), not a value comparison.

The rerun has two remaining reds, and they are the same event:

job	step the job fails at	what it says
chat-bundle-browser (2)	Qualify the actual compiled UI	Error: locator.waitFor: Timeout 30000ms exceeded.
checks	Require the planned backend and Dashboard qualification	ValueError: chat-bundle-browser must be success

checks is the aggregator that re-asserts the shard above; it is not an independent failure. The sibling shards chat-bundle-browser (1) and (3) pass, as do all three dashboard-browser shards, all three typescript-core shards and every stage2c variant.

chat-bundle-browser (2) failed at this same step, with a 30 s Playwright wait, on the previous PR from this fork (#7539) — which changed a docstring and one control-plane test. So the shard is unstable independently of anything in this branch.

Scope reminder: this PR changes loopx/presentation/projections/session_dash.py and adds tests/presentation/test_session_dash_status_bucket.py. It touches nothing in loopx/web/chat/ or the dashboard bundle.

test-shard (1..6) never left pending on either run of this head, so those shards have not reported at all.

[loopx-agent]

Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; reasoning_effort=max.

动机 / Motivation

查看会话看板的操作者需要分清正在执行、等待本人、阻塞和已完成的目标。原先 installed 会因含有 stalled 被标为阻塞，incomplete 与 not_completed 会误报完成，waiting_on=controller 则遗漏人工等待；新版本修正这些徽标和总数。这些误报使操作者把未完成工作当作结束，或把健康状态当成阻塞。同一份24目标输入经真实导出CLI和HTTP读取，原版7项误分类在新版本全部修正，其余17项及全部非分类字段保持相同。本提案只修正既有只读看板的分类，不改变任务执行、权限、调度或持久状态。

改动思路 / Approach

沿用 loopx dash generate 与 DashRequestHandler 到同一个 build_session_dash_projection 的路径，只在展示层解释现有状态。匹配改为有边界的标记，完成信号先排除既有否定前缀，并消费运行时已经发出的 controller 等待值；完成、阻塞、人工等待、活动的优先级保持。无变化与更正后的信息仍由原HTML渲染器显示。

架构判断：保留既有投影边界，两个局部匹配函数足以修正已复现错误；无需增加状态库、共享协议或第二个权限判断。当前PR边界：仅状态桶、汇总计数及针对真实投影的回归保护。对已完成目标，不因 waiting_on=controller 又要求操作者介入；真实 blocked 也仍先于人工等待。

相同输入	原分类	当前分类
installed / uninstalled	blocked	other
incomplete / undone / not_completed / not_shipped	done	other
未结束且 waiting_on=controller	other	needs_user

具体改动 / Changes

准确审查 64975e2f98b35c4eff429f1e29d3f172adda2308 的全部两个路径：生产代码+48/-4，测试+120；PR差分相对共同祖先 5ebeb55796883284db82f06c7f0dcfed9bf9b65f，没有把分支落后的主干历史算进提案。真实合并树检查可与已观察主干干净组合。

关键代码讲解

• _has_status_marker:80 对转义后的标记使用字母数字边界，阻止 installed、incomplete 等词内命中。
• _status_bucket:129 在既有优先级中处理否定完成和 controller；它只返回展示桶，不写控制面。
• _goal_entry:189 继续读取 attention/run 投影，保留原始状态、等待原因、Todo与最近执行；汇总复用这一结果，避免徽标与统计各自解释。

规范依据 docs/product/surfaces/session-dash-panel-design.md，固定版本 5ebeb55796883284db82f06c7f0dcfed9bf9b65f：原接受文本。该文没有编号，以下直接使用原小节/条目名：Overview、Per goal 已由桶和徽标对照实现；Command Surface 的只读CLI/HTTP、Public/Private Boundary 的转义/导出扫描保持，现有公开smoke通过。

对主干的风险 / Risks

PASS：全 tests/presentation 50项、原 session-dash smoke、两文件Ruff，以及标准premerge的5个直接检查和5个选中检查；严格CQR独立核验有效。24目标实路对照使用合成 collect_status 输入，真实执行CLI解析、处理器、HTTP、投影、HTML和边界扫描；两次重复、状态更正再恢复、来源字节不变均通过。原比较投影与共同祖先逐字相同，但周边CLI并非把整个旧仓库重建，未据此声称全主干集成或安装态覆盖。

首次把旧分支验证器接到较新的操作环境时，检查在读取不支持的 context_delivery 配置命名空间处退出。保留该失败；随后使用源分支本地premerge，并由当前验证器单独强制同一准确作用域的严格CQR，未修改任何共享配置、断言或门槛。没有查询远端CI；作者关于旧CI的说明不作为本次独立归因。

剩余风险是自由文本分类仍属兼容启发式，例如既有 not_failed 仍可能被当成 blocked；当前修复未解决所有自然语言否定。未来有规范生产者字段时，应在既有投影owner优先消费显式类型、保留有界旧输入映射；这不构成本次已验证错误修复的阻断。新词组仅为本地展示匹配，不是权限或共享状态枚举。

我的整体评价 / Verdict

未发现阻断问题，APPROVE。长程工作语义 preserved：反复读取和更正不写任务或产生执行效果；用户体验 improved：人工等待和完成总数与具体目标相符，已有意图无需新增确认或配置。视觉观察：实际HTML使用3会话、24个长中英标题及混合状态；检查1200桌面整页、390首屏和下方完整长行，徽标与主读数更正，既有窄屏横向表格仍需横滚，右侧等待/行动完整可读。当前页面没有折叠/展开控制，默认即完整行；下方和横向内容已查看。未验证范围：正式安装App、自然现场新鲜度、完整无障碍与所有自由文本语义；这些不被合成画面或无溢出断言替代。此次代码规模与错误影响相称，局部词匹配的解释边界明确。合并资格及授权另行判断。

English verdict: APPROVE — the exact-head read-only classification fix is verified; no blocking finding.