Add Volcengine smoke and full Provider replay

[2002yy]

Scope

Continue P0-E3 with a replay-only Volcengine Ark adapter and bounded smoke/full workflow modes. This measures compatibility before any production Provider expansion.

Changes

• add VolcengineArkReplayProvider as a subclass of the provenance-approved real-provider replay adapter
• use the official OpenAI-compatible Ark base URL https://ark.cn-beijing.volces.com/api/v3
• accept VOLCENGINE_API_KEY with ARK_API_KEY fallback and an exact Model ID / Endpoint ID
• keep Volcengine out of the production chat Provider owner
• add explicit --model-name and --base-url replay arguments
• add a default smoke mode that runs only clean_requests_session
• keep AnswerClaim evaluation exclusive to full, which runs all 10 gold cases
• require provider identity, model name, endpoint fingerprint, exact smoke/full scope and record-only boundaries
• reject Agent Plan /api/plan/v3 and Coding Plan /api/coding/v3 in adapter code, tests and documentation
• add adapter, routing, identity, JSON mode, usage, scope and workflow regression tests

Boundaries

• no production ChatTurn or SQLite write
• no committed learning truth change
• no claim UI or production claim producer
• Volcengine is replay-only and does not appear in the production chat Provider owner
• no automatic Provider spend in PR CI
• smoke is connectivity/schema evidence only and cannot produce a model-quality conclusion
• a real workflow still requires a configured GitHub secret and manual dispatch

Validation

Final head CI #1696 passed the complete repository gate: pytest, RAG K1, Ruff, packaging, detect-secrets, expanded mypy, frontend tests/build, 35/35 fixture+narrow browser journeys and 12/12 real-stack cases.

No real Volcengine request was made. After merge, the next operation is one manually dispatched Volcengine smoke using the exact Model/Endpoint ID. Only a green smoke may proceed to two full 10-case runs.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 898cb0599e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Reject smoke replays with unparsable provider output

When Volcengine returns plain text or malformed JSON, parse_provider_answer() records a fallback candidate with a nonempty results[].parse_error, while run_provider_answer_replay() still reports status: completed and produces an answer_quality dictionary. This smoke branch checks only counts and scope, so that response passes the workflow even though the smoke is intended to establish JSON/schema compatibility before a full paid run; inspect the selected result's parse error and fail the smoke when it is nonempty.

Useful? React with 👍 / 👎.