provider AKI.IO: update models 2 added, 1 removed

[aki-io-team]

New models added:

• deepseek v4 flash 0731
• qwen 3.8 27b

Models removed:

• minimax m2.5

[github-actions]

Action items

• [high] [possible mistake] providers/aki-io/models/deepseek-v4-flash-0731-284b.toml:5 - Check: Relay reasoning_options must match the lab/same-surface baseline for this model (toggle or equivalent off control + Flash effort). Why: First-party DeepSeek V4 Flash and established relays use toggle + effort low/high/max (or none + those levels when off is effort-mapped). This entry only lists low/high/max, so callers cannot disable reasoning even though the lab exposes off and other AKI.IO models document none for that. Action: Align with this host’s real wire surface: either toggle + ["low", "high", "max"] (with a leading toggle wire comment) or ["none", "low", "high", "max"] if AKI.IO disables via none; do not leave graded effort without an off path unless you can show the API has no disable control.
• [medium] [violation] providers/aki-io/models/deepseek-v4-flash-0731-284b.toml:3 - Check: After base_model, only provider-specific fields and real deltas. Why: attachment = false and reasoning = true match models/deepseek/deepseek-v4-flash-0731.toml unchanged, so they are redundant restates. Action: Drop attachment and reasoning; keep cost, reasoning_options, and any true limit/modality deltas.
• [medium] [violation] providers/aki-io/models/qwen3.8-27b.toml:5 - Check: Override-only provider files after base_model. Why: reasoning = true is already true on alibaba/qwen3.8-27b and adds no host-specific delta (attachment = false and narrowed modalities are fine). Action: Remove the redundant reasoning = true line.
• [low] [possible mistake] .pr-review/pull-request.json - Check: Data-changing PRs should cite first-party pricing/docs/API and map each citation to claims. Why: New DeepSeek/Qwen costs (and limits), plus Kimi cache_read, cannot be reviewed from the PR body alone. Action: Add direct AKI.IO (or other first-party) links and note what each supports (list prices, cache rate, context/output caps).

[github-actions]

No actionable findings.