fix(smokes): track the dsh runner's reasoning_effort keyword

[huangruiteng]

What changed

main's full-public-smokes workflow is red, and the failure is not in the
host: two hermetic fake-runner fixtures still declare the run_dsh_turn
keyword set the adapter used before the managed execution profile landed. The
adapter has supplied reasoning_effort since then, so the fake runner raises

TypeError: run_dsh_turn() got an unexpected keyword argument 'reasoning_effort'

which the adapter maps to BuiltInHostError("dsh_execution_failed") and the
Turn reports as an unattributed host_failure. The observable symptom is
examples/loopx-turn-dsh-builtin-host-e2e-smoke.py failing its success leg
(exit 1, result_kind: host_failure), followed by smoke-fleet-health failing
its .ready == true check.

This is pre-existing on main, not fallout from the recent steward/managed
work: the same smoke fails at 001e0c380, before the credential-resolved Turn
default and the chat PRs landed.

Changed surfaces

Public example smokes only; no runtime, CLI, schema, or frontend change.

• examples/loopx-turn-dsh-e2e-smoke.py: the fake runner source is now emitted
  from one module-level FAKE_DSH_RUNNER_SIGNATURE, carrying the current
  keyword set (... provider, model, reasoning_effort, max_tokens, ...) and
  asserting the resolved provider/model/effort it is handed.
• examples/dsh-turn-host-adapter-smoke.py: loads that smoke the same way the
  built-in host smoke already does and uses the same emitted source instead of
  keeping a second hand-written copy of the signature.

The built-in host smoke needs no change: it already reuses the generic-cli
smoke's writer, which is why fixing the writer fixes both failing legs.

Why one source

Three smokes load a fake runner and the contract is real (the adapter documents
the keyword set on load_dsh_runner). A fixture that lags it fails as an
unattributed host_failure at the endpoint instead of naming the seam that
moved, so the emitted signature now lives in the module the other host smokes
already import: the next adapter keyword lands once.

Validation

check	result
python examples/dsh-turn-host-adapter-smoke.py	11 checks passed
python examples/loopx-turn-dsh-e2e-smoke.py	passed
python examples/loopx-turn-dsh-builtin-host-e2e-smoke.py	passed (was failing: host_failure, exit 1)
python -m pytest tests/test_dsh_goal_mode.py -q	58 passed
python examples/loopx-turn-managed-step-self-heal-smoke.py	passed
python -m ruff check on both changed files	passed
python scripts/generate_semantic_inventory.py --check	up to date
loopx canary premerge --from-git-diff	gate passed, 0 failures, 0 manual holds, 3 selected checks
change-quality receipt	cqr_0ea46e303db56e9978c1 at scope 0ea46e303db56e99

Negative path: removing reasoning_effort from the shared signature makes
examples/dsh-turn-host-adapter-smoke.py fail with exactly the CI
TypeError; the mutation was reverted and the smoke passes again. The repaired
fixture is load-bearing rather than merely green.

Boundary and remaining limits

• Two other hermetic runners in this family accept **kwargs
  (examples/loopx-turn-managed-step-self-heal-smoke.py, the
  tests/*_runner.py helpers), so they cannot catch this drift class by
  construction. They are left as they are: tests/dsh_goal_mode_fake_runner.py
  plus the two smokes repaired here are the exact-keyword guards.
• No DeepSeek Harness SDK, credential, network, or model call is involved: the
  whole path is hermetic and deterministic.
• ruff format reports both files as non-conforming on main as well; that
  pre-existing formatting drift is not folded into this fix.

[huangruiteng]

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

动机

main 上的全量公开冒烟是红的，但红的不是 host，而是两份跟 host 契约脱节的 hermetic fake runner：adapter 在托管执行档落地后一直传 reasoning_effort，两个 fake runner 却还写着旧的关键字集，于是调用点抛 TypeError: run_dsh_turn() got an unexpected keyword argument 'reasoning_effort'，被 adapter 映射成 BuiltInHostError("dsh_execution_failed")，Turn 只能报一个不带归因的 host_failure。结果是 examples/loopx-turn-dsh-builtin-host-e2e-smoke.py 的成功腿挂掉、smoke-fleet-health 跟着 .ready == true 失败。这条修复把"假 runner 必须跟上真实调用契约"这件事重新做成可执行的门，并且顺带消掉一份重复的签名。

改动思路

1. fake runner 的签名只保留一份：放在 examples/loopx-turn-dsh-e2e-smoke.py 的 FAKE_DSH_RUNNER_SIGNATURE，因为内置 host 冒烟本来就 import 这个模块复用它；examples/dsh-turn-host-adapter-smoke.py 改成同一份来源，不再手写第二份。
2. 签名按 adapter 在 load_dsh_runner 文档里声明的关键字集写全，并要求 runner 断言拿到已解析的 provider/model/reasoning_effort——profile 被丢掉时在 seam 上失败，而不是到真实端点才失败。
3. 不改 **kwargs 的 runner（examples/loopx-turn-managed-step-self-heal-smoke.py 与 tests/*_runner.py）：它们服务于各自更窄的断言，精确关键字的守卫由本 PR 修的两份冒烟加 tests/dsh_goal_mode_fake_runner.py 承担。

具体改动

• examples/loopx-turn-dsh-e2e-smoke.py（+14/-5）：新增模块级 FAKE_DSH_RUNNER_SIGNATURE（含 reasoning_effort 与 profile 断言），_write_fake_dsh_runner 改用它生成代码。
• examples/dsh-turn-host-adapter-smoke.py（+14/-4）：按内置 host 冒烟同样的 importlib 方式加载通用 e2e 冒烟模块，runner 文本用同一份签名。
• 未改任何产品代码、CLI、schema、前端；内置 host 冒烟本身无需改动，因为它复用的就是修好的 writer。

关键内容讲解

1. 失败是被证伪出来的，不是推断出来的：先在干净 worktree 上跑出 host_failure/exit 1，再在 001e0c380（更早的 main）上复现同样失败，确认这不是最近几个管家/managed PR 带来的回归。
2. 修复方向由"哪份契约为准"决定：adapter 侧 load_dsh_runner 的 docstring 与真实调用点已经声明关键字集，所以改的是 fixture，不是产品。
3. 负向验证：把 reasoning_effort 从共享签名里拿掉，examples/dsh-turn-host-adapter-smoke.py 立刻以 CI 里那句同样的 TypeError 失败；改回后重新通过。因此这份 fixture 是承重的，不是"碰巧绿"。

对主干的风险

P3（可接受）：三份冒烟现在共享一个模块级常量，examples/dsh-turn-host-adapter-smoke.py 因此在导入期多加载一个 example 模块（无副作用，只定义常量与函数）。代价是一次导入，收益是签名不会再各自漂移。

P3（已知残余）：接受 **kwargs 的 runner 按构造就无法捕获这一类漂移，本 PR 没有扩大它们的断言范围；精确关键字守卫落在两份被修的冒烟与 tests/dsh_goal_mode_fake_runner.py 上。

其余残余：变更只动公开示例冒烟，loopx canary premerge --from-git-diff 的门通过（0 失败、0 人工挂起、3 项选中检查含 public-boundary 扫描与语义漂移冒烟）；本地验证覆盖 adapter 往返、通用 e2e 全链路、内置 host 两条腿、tests/test_dsh_goal_mode.py 58 项、ruff、语义清单检查。未覆盖：真实 DeepSeek Harness SDK/端点的调用（本路径本来就是 hermetic，不涉及凭据与网络）。

我的整体评价

APPROVE。这是一个归因清楚、边界最小、并且带负向证据的 CI 修复：fixture 跟上真实契约、签名单一来源、失败信息可读。它同时让 main 的公开冒烟舰队恢复绿色，属于可以直接自合并的类型。变更范围仅两个公开示例文件，change-quality receipt cqr_0ea46e303db56e9978c1 覆盖同一 scope fingerprint。

English verdict: APPROVE at 5efef15. main's full-public-smokes is red because two hermetic fake-runner fixtures still declare the dsh keyword set as it was before the managed execution profile landed; the adapter supplies reasoning_effort, the fake runner raises TypeError, and the Turn degrades to an unattributed host_failure. Reproduced on the worktree and again at 001e0c380, so it is not fallout from the recent steward/managed PRs. The fix emits the fake runner signature from one place (the generic-cli e2e smoke the built-in host smoke already imports) and has the emitted runner assert the resolved provider/model/effort, so a dropped profile fails at the seam. Verified: adapter smoke 11 checks, generic e2e, built-in host e2e (previously exit 1 / host_failure), 58 goal-mode tests, ruff, semantic inventory check, canary premerge gate passed with 0 failures, and a mutation that removes reasoning_effort reproduces the exact CI TypeError before being reverted. Scope is two public example smokes; no runtime, CLI, schema, or frontend change. Receipt cqr_0ea46e303db56e9978c1.