Polish progressive Recompose demo behavior

[Phloraxx]

Why

Post-merge audit of the Recompose/voice release found two small demo-quality issues:

1. three judge presets were implicitly forcing reduceMotion=true, which hid the progressive View Transition choreography even when the person's actual profile allowed motion;
2. when Ollama/Qwen was unavailable, the deterministic Recompose surface stayed active correctly, but the local result was reported as applied=true, causing the shell to claim that Qwen had successfully personalized the page.

Changes

• preserve the person's real reduced-motion preference across all judge presets instead of silently overriding it
• keep progressive Recompose motion visible for motion-enabled judge demos while still honoring Learn Me/system reduced-motion behavior
• report local Qwen failure as a local fallback while leaving the deterministic Recompose surface active
• extend preset tests to prove reduced-motion preservation for both motion-enabled and reduced-motion profiles
• correct the Recompose/voice source-of-truth wording to match the implemented auto-submit dictation flow and fallback semantics

No new product scope. This is release/demo truthfulness and interaction polish only.