[Air000000]

Outcome: NOT release-accepted

The complete offline suite passes, but the separately authorized real-provider ask acceptance stopped on its first positive case with HTTP 503. This PR records the failure, not a successful rollout or a model-quality promotion. No business/runtime algorithm, prompt, timeout, default switch or ticket path is changed.

Offline coverage

• Installed both already-declared requirement files in the isolated virtual environment, resolving initial missing optional dependencies / 22 collection-error files.
• Entire tests/ on Windows Python 3.12.4: latest 622 passed, zero failures/skips. Earlier pre-guard run: 621 passed with one third-party Numba cast warning.
• Main pytest process denies external sockets and private project data/ access; fixtures use temporary databases/data. HF offline mode and DeepEval telemetry opt-outs enabled.
• Extend CI beyond the old focused subset: install declared evaluation dependencies, then run the complete offline directory under the same safety guard.
• Add acceptance transport tests for ground-truth/extra-field/foreign-body rejection and the six-call classifier ceiling. Ruff, pip dependency consistency, workflow YAML/inline-Python compilation and diff checks pass.

Bounded ask acceptance

Reuse the existing plan-only smoke entry point with --ask; add a fixed seven-document/eight-case fixture module. Expected outcomes remain local and are not sent to models. Default search smoke retains its old 8/3 caps.

Human approved at most 14 Beijing text-embedding-v4, 6 Singapore qwen3-rerank, 6 Singapore qwen3.5-plus-2026-04-20 classifiers and 2 Beijing qwen3.5-plus generations, planning budget CNY 2. No automatic retry/redirect; abort on the first unexpected result. Endpoint/model/text/prompt/body-size allowlists are checked before sending.

Observed once, then stopped

• Actually attempted: 9 embedding + 1 rerank + 1 classifier + 1 generation = 12.
• Embedding/rerank/classifier returned HTTP 200; known usage totals 2639 tokens.
• Generation produced no usable response within the 30-second timeout; transport duration 30,059 ms, HTTP status and usage unknown. It may have been billed. API returned 503.
• First positive case failed; the remaining seven real cases were NOT RUN. No paid rerun, no increased timeout, no post-hoc prompt changes.
• A semantic warning is preserved too: the classifier marked version_and_time=NOT_REQUIRED despite an explicit version 1.0 constraint. The source covers 1.0, but this audit status contradicts the policy; a valid schema/locator is not semantic proof.
• Generation omits enable_thinking, while official docs say Qwen3.5 Plus thinking is enabled by default. This is a configuration risk to investigate, not a proven unique cause of this timeout. Exact timeout phase was not recorded; do not invent one.

Safe per-attempt metadata, classifier completion, case statuses and execution-time script hashes are retained in docs/reports/hybrid_acceptance_2026-10-17.json.

Still pending

Generation configuration/timeout diagnosis and an explicitly authorized re-test; remaining real cases; independent semantic acceptance. Docker engine is unavailable, so live-server HTTP/restart/persistence/rollback acceptance is not passed. Ticket Agent remains Dense; no all-Hybrid chain or public deployment claim.

Primary checkout changes, existing databases, .env and persistent switches remain untouched. No business documents, TechQA, ground-truth labels or secrets in model inputs; credentials only used for normal provider authentication. Merging this PR does not enable the experiment or resolve the observed failure.