fix(vision): route standalone vision through native providers

[huangzesen]

Summary

• register codex-pool and codex_pool as native Codex vision aliases and route the Codex family through the existing Responses input_image service
• preserve same-provider identity for direct-native OpenAI, Anthropic, and Gemini vision: explicit capability overrides win; otherwise the active model is reused, and OpenAI/Anthropic also reuse the active endpoint
• keep MiMo on an explicit or known vision-capable model (mimo-v2.5 by default) while preserving its same-provider endpoint; do not forward text-only MiMo chat models
• map glm to the existing Zhipu MCP vision service; keep MiniMax/Zhipu MCP routing unchanged
• keep OpenRouter, DeepSeek, Kimi, custom relays, Claude Code aliases, and local vision unavailable by default unless the user explicitly configures the existing compatibility/skill+MCP path
• add focused routing/isolation/request regressions and update the vision Contract/Anatomy documents

Scope

This PR keeps the existing explicit vision(image_path, question) tool. It does not add canonical main-session image blocks, auto-enable omitted optional capabilities, infer whether a relay/upstream model accepts images, install/download a local model, or introduce any silent cross-provider fallback.

MiniMax and Zhipu remain explicitly configured MCP-backed services. Built-in preset declarations/editor choices are owned by the TUI repository; this kernel PR makes declared vision capabilities route correctly but does not silently turn on vision for a TUI preset that omits it (notably the current Gemini preset source).

Validation

Parent-run validation on exact head d8abcdc772e88228ce5efcf07f6c3526f8f1fd11:

• /Users/huangzesen/work/GitHub/lingtai-kernel/.venv/bin/python -m pytest tests/test_vision_capability.py tests/test_vision_services.py -q — 56 passed
• /Users/huangzesen/work/GitHub/lingtai-kernel/.venv/bin/python -m pytest tests/test_codex_pool*.py -q — 109 passed
• /Users/huangzesen/work/GitHub/lingtai-kernel/.venv/bin/python -m pytest tests/test_architecture_documents.py tests/test_anatomy_drift_checker.py -q — 12 passed
• /Users/huangzesen/work/GitHub/lingtai-kernel/.venv/bin/python -m lingtai.tools.glossary_validator --check — 54 resources validated
• py_compile, lazy-import smoke (lingtai.tools.vision does not import lingtai.services), and git diff --check — passed

Live evidence and safety

After the initial Codex production commit 0a9fdfc66f004068477cc6394812ac8c4e09e598, one explicitly authorized controlled PNG smoke through the active codex-pool / gpt-5.6-sol route returned the exact red/yellow/white/blue geometry. The broader provider expansion was validated only with mocked/local tests: there were no live OpenAI, Anthropic, Gemini, MiMo, MiniMax, or Zhipu calls.

The current live runtime remains on 0a9fdfc6; this expanded PR head was not installed or refreshed. No auth/config mutation, credential-file inspection, fallback-provider call, dependency change, or local model download was performed.

[huangzesen]

Superseded by #964, which carries the latest-main implementation and the complete current provider/preset coverage for Lingtai-AI/lingtai#957. Thank you for establishing the original Codex/Codex-pool vision direction here; that work is preserved and extended in the replacement.