Three fixes for flush_memories / compression context window overflow:
1. ALWAYS deduct headroom before comparing aux_context vs threshold.
#15631 only deducted inside 'if aux_context < threshold' — which
never fires in the common same-model case (threshold = context × 0.50
means aux_context > threshold always). Now headroom is computed
unconditionally and effective_limit = aux_context - headroom is
compared against threshold.
2. Also resolve flush_memories auxiliary model in the feasibility check.
If the user configures separate auxiliary.flush_memories provider,
the flush model's smaller context was unchecked.
3. Defence-in-depth trimming in flush_memories() for CLI /new and
gateway resets that bypass preflight compression entirely.
25d9fc8094