fix: prefer native vision in auto image routing

[cosmopolitan033]

Summary

• In agent.image_input_mode: auto, prefer native image attachments when the active main model is known to support vision.
• Keep explicit auxiliary.vision as the fallback path for text-only or unknown main models.
• Update image-routing tests to cover the mixed-profile case described in the issue.

Fixes #29135

Test plan

• /Users/cosmopolitan/.hermes/hermes-agent/venv/bin/python -m pytest -o 'addopts=' tests/agent/test_image_routing.py -q
• git diff --check

[cosmopolitan033]

Maintenance update: I rebased/merged this PR locally against current main and resolved the conflict in agent/image_routing.py by preserving the current config-override-aware capability lookup while keeping the PR behavior: native-capable main models use native image parts before falling back to an explicit auxiliary vision backend for text-only/unknown models.\n\nValidation run locally after resolving conflicts:\n\nconsole\n$ .venv/bin/python -m pytest tests/agent/test_image_routing.py\n58 passed in 0.51s\n\n\nI attempted to push the conflict-resolution commit to the PR branch, but GitHub rejected the push because this token lacks the workflow scope and the fork needs to carry upstream workflow changes (remote error: refusing to allow an OAuth App to create or update workflow .github/workflows/docker-publish.yml without workflow scope). I will retry updating the branch when a workflow-scoped credential or maintainer-side branch update is available.