fix(chat): detect bare-named Qwen3.6 tags as vision-capable

[SumTuusDeus]

Summary

is_vision_model() returns False for every Qwen3.6 tag (qwen3.6, qwen3.6:27b, qwen3.6-35b-a3b, fine-tune names containing the family name), so Odysseus silently swaps image attachments for text captions and the model never sees the picture — the exact failure mode #1274 fixed for gemma3:4b-style tags. Unlike Qwen2.5/Qwen3, the Qwen3.6 family ships vision natively in every bare-named checkpoint and has no separate VL variant (all official repos are image-text-to-text with a built-in vision encoder; vLLM serves them with vision on by default — full evidence with links in #5501), so none of the existing qwen*-vl keywords or the standalone-vl regex can ever match. This PR adds the single bare family keyword "qwen3.6" to _VISION_MODEL_KEYWORDS with a justifying comment matching the gemma3/llama4/phi-4 entries, per the err-toward-True policy (#124), plus test coverage. It complements open PR #3743 (Ollama /api/show probe): the probe can't help vLLM/OpenAI-compatible endpoints, where the keyword list remains the only decision path, and this also fixes the fallback list that probe keeps relying on.

Per the contribution policy on auto-generated PRs: this is an agent-assisted, human-reviewed contribution; issue #5501 was filed first describing the problem, before this PR.

Target branch

• This PR targets dev, not main.

Linked Issue

Fixes #5501
Related: #124, #1274, #4723 (same symptom reported for qwen 3.6 on Ollama), #3743

Type of Change

• Bug fix (non-breaking — fixes a confirmed issue)
• New feature (non-breaking — adds new behaviour)
• Breaking change (changes or removes existing behaviour)
• Refactor / cleanup (behaviour unchanged)
• Documentation only
• CI / tooling / configuration

Checklist

• I searched open issues and open PRs — this is not a duplicate. Closest neighbours are linked above: Images cannot be seen by model that is vision capable #4723 (symptom report, Ollama/Windows) and fix(vision): probe Ollama /api/show capabilities instead of guessing … #3743 (Ollama-only probe; leaves the keyword list stale for other backends).
• This PR targets dev
• My changes are limited to the scope described above — one keyword + its comment + tests; no unrelated changes.
• I actually ran the app and verified the change works end-to-end: on a live Docker install (docker compose up -d --build) pointed at a vLLM-served Qwen3.6-35B fine-tune, image attachments reach the model natively with this keyword present, instead of being caption-swapped; verified is_vision_model("qwen3.6-35b") is True inside the running container.

How to Test

1. Quick unit check on this branch: python -m pytest tests/test_vision_model_detection.py (5 pass; new cases cover qwen3.6, qwen3.6:27b, qwen3.6-35b-a3b positive and qwen3:8b, qwen2.5:14b negative). Also ran python -m pytest tests/test_lmstudio_vision.py (10 pass) and python -m py_compile src/chat_helpers.py tests/test_vision_model_detection.py.
2. In a REPL on this branch: from src.chat_helpers import is_vision_model; is_vision_model("qwen3.6:27b") → True (was False on dev).
3. End-to-end: run Odysseus against any OpenAI-compatible endpoint serving a Qwen3.6 checkpoint (vLLM recipe: https://recipes.vllm.ai/Qwen/Qwen3.6-27B), attach an image, and ask about it — the model now describes the actual image instead of a caption.
4. Regression: text-only tags (qwen3:8b, qwen2.5:14b, mistral-small, deepseek-r1) are still not flagged as vision.

Visual / UI changes — REQUIRED if you touched anything that renders

Not a visual change — no UI/rendering paths touched.

• I am not an LLM agent submitting a bulk PR. Single scoped fix; agent-assisted and human-reviewed, with issue is_vision_model() misses the entire Qwen3.6 family — images silently caption-swapped on vLLM/OpenAI-compatible endpoints #5501 filed first per the policy above.

[AbdullahWins]

Looks good to me! The logic for detecting bare-named tags as vision-capable seems correct and the accompanying test coverage is appreciated. LGTM.