fix(amazon-bedrock): correct hosted reasoner catalog

[rekram1-node]

Summary

xAI/NVIDIA/MiniMax/Z.AI/Moonshot split of #6794, based on current dev.

• Correct verified dates, limits, capabilities, prices, and host reasoning controls across 16 entries.
• Mark models interleaved where Bedrock Converse returns reasoningContent.
• Confirm MiniMax M2/M2.1/M2.5 structured output and reasoning side channels through live Converse checks.
• Retain Kimi K2.5’s temperature override after live acceptance of both temperature=0 and temperature=1.
• Keep unresolved NVIDIA caller controls conservative rather than inventing options.

Sources and live-check notes are in leading TOML comments.

Validation

• bun validate
• git diff --check
• No removals or route changes

[github-actions]

Action items

• [medium] [possible mistake] providers/amazon-bedrock/models/nvidia.nemotron-nano-3-30b.toml:15 - Check: Provider limit overrides must stay host-accurate after base_model merge (plain objects deep-merge; omitted keys inherit the lab value). Why: The diff drops the existing limit.context = 128_000 override and only keeps output = 8_192. The resolved context therefore becomes the lab value 262_144 from models/nvidia/nemotron-3-nano-30b-a3b.toml, roughly doubling the prior Bedrock figure. The leading live-check note only documents maxTokens=8192; it does not support a 262k Bedrock context window. Action: Either restore a verified Bedrock limit.context override, or cite host evidence that this route actually serves 262k context and keep the inherited lab value intentionally.

[rekram1-node]

Addressed in the latest commit. AWS’s current Nemotron Nano 3 card explicitly documents a 256K context window and 8K max output for this Bedrock route. The inherited lab context 262_144 is therefore intentional and correct; the prior 128K provider override was stale. Added a leading note making that inheritance explicit. bun validate and git diff --check pass.

[github-actions]

No actionable findings.