fix(vision): describe images via aux vision on text-only model failover

[viz-A-viz]

Problem

When the primary model (vision-capable) dies mid-turn and the fallback chain lands on a text-only model (e.g. DeepSeek), two bugs occur:

1. Inline image parts in api_messages cause 400 errors — the text-only model rejects image_url parts it cannot process.

2. vision_analyze tool still uses the native fast path — _should_use_native_vision_fast_path() reads stale runtime-main globals (still pointing at the vision-capable primary), so every vision_analyze call returns image bytes that the text-only fallback model cannot process.

Fix

_sync_failover_vision_support (called after every provider failover):

• Looks up the failover model's vision capability via _lookup_supports_vision
• If text-only and image parts present — describes each image via the auxiliary vision LLM and replaces the part in place with a text description
• Falls back to a plaintext placeholder if aux vision fails
• Sets agent._vision_supported = False
• Sets module-level _failover_vision_disabled flag in vision_tools

_should_use_native_vision_fast_path (vision_tools.py):

• Checks _failover_vision_disabled at the top — returns False when set, forcing vision_analyze through the auxiliary vision text path

turn_context.build_turn_context:

• Clears _failover_vision_disabled on primary restore, re-enabling the native fast path when the vision-capable primary is back

Tests

TestFailoverVisionSupport (5 cases): text-only replacement, aux-vision failure placeholder, vision-capable pass-through, already-unsupported skip, tool-message linkage.

TestFailoverVisionDisablesFastPath (2 cases): text-only failover sets the vision_tools flag, fast path returns False when flag is set.

18 passed in 0.64s