fix(amazon-bedrock): correct open model catalog

[rekram1-node]

Summary

Remaining open-model split of #6637, based on current dev: DeepSeek, Qwen, xAI, NVIDIA, Google, MiniMax, Z.AI, and Moonshot.

• Audit all 30 existing entries; preserve every existing ID and route.
• Correct confirmed prices, dates, modalities, open-weight facts, limits, and host reasoning controls.
• Convert third-party provider definitions to override-only base_model entries.
• Add three live-verified Gemma 4 Mantle Responses models and complete Gemma 3 lab metadata.
• Keep Gemma 3 tool_call=false: forced calls can be emitted, but valid native tool-result continuations fail.
• Keep Bedrock DeepSeek R1 tool_call=false after explicit API rejection.
• Retain Kimi K2.5’s temperature override after live Converse accepted both temperature=0 and temperature=1.

Unresolved NVIDIA controls remain conservative rather than inventing caller options. New Chat/GovCloud candidates with incomplete controls or client-route support are deferred.

Sources are in leading TOML comments, including AWS model cards and public pricing.

Validation

• bun validate
• All new lab entries contain required capabilities, modalities, context, and output
• git diff --check
• No model deletions or route changes

[github-actions]

Action items

• [high] [violation] models/google/gemma-3-4b-it.toml:11 - Check: Lab metadata must stay provider-agnostic; host-only capability limits belong on the provider entry. Why: tool_call = false is justified only by Bedrock Converse tool-result failures (PR body + Bedrock comments). Lab files are shared facts about the model; this value will be inherited by any future base_model consumer and contradicts other hosts that already advertise Gemma 3 tool use. Bedrock R1 correctly keeps lab tool_call = true and overrides on the provider. Action: Set lab Gemma 3 IT entries to the model-level capability (likely tool_call = true), and set tool_call = false on providers/amazon-bedrock/models/google.gemma-3-{4b,12b,27b}-it.toml as the host-specific override.
• [medium] [possible mistake] providers/amazon-bedrock/models/deepseek.v3-v1:0.toml:7 - Check: Provider models that return a reasoning side channel should set interleaved for this API. Why: Leading comments for V3.1 and V3.2 state live Converse returns reasoningContent when thinking is on, but neither file sets interleaved, while the same PR sets interleaved = true for Kimi/GLM peers that use the same side channel. Action: Verify Converse reasoning content shape for deepseek.v3-v1:0 and deepseek.v3.2; add interleaved = true (or field = "reasoning_content") if reasoningContent is returned, or correct the comments if it is not.
• [medium] [possible mistake] providers/amazon-bedrock/models/zai.glm-4.7-flash.toml:7 - Check: Same-host GLM reasoning peers should mark the reasoning side channel consistently. Why: Live notes say reasoning_config=high returns reasoningContent, and zai.glm-4.7 / zai.glm-5 in this PR set interleaved = true, but flash omits it. Action: Confirm whether flash returns the same reasoning side channel; if yes, add interleaved to match the sibling GLM Bedrock entries.

[rekram1-node]

Addressed in 38611e3.

• Gemma 3 lab metadata now records the provider-agnostic model capability with tool_call = true for 4B/12B/27B.
• Each Bedrock Gemma 3 provider entry explicitly overrides tool_call = false, preserving the live Bedrock finding that forced tool generation works but native tool-result continuation fails.
• Added interleaved = true to DeepSeek V3.1, DeepSeek V3.2, and GLM 4.7 Flash. Live Converse evidence already recorded in the headers confirms that thinking mode returns reasoningContent; default/off checks omit it.

bun validate and git diff --check pass.

[github-actions]

Action items

• [high] [possible mistake] providers/amazon-bedrock/models/minimax.minimax-m2.5.toml:4 - Check: Provider interleaved must be set when this host returns a reasoning side channel. Why: The live note says Converse returns reasoningContent for MiniMax M2.5 (and M2/M2.1 share the same always-on Bedrock surface), but none of the MiniMax entries set interleaved, while sibling Bedrock reasoners that emit reasoningContent (DeepSeek V3.x, GLM, Kimi) do. Clients will miss the reasoning channel. Action: Add interleaved = true (or the correct field form) on minimax.minimax-m2.toml, minimax.minimax-m2.1.toml, and minimax.minimax-m2.5.toml, or document why MiniMax does not expose a parseable side channel on this host.
• [high] [possible mistake] providers/amazon-bedrock/models/qwen.qwen3-32b-v1:0.toml:7 - Check: Provider interleaved must be set when this host returns a reasoning side channel. Why: The live note states reasoning_config=high returns reasoningContent, but the file only sets reasoning_options = [{ type = "toggle" }] and omits interleaved, unlike other Bedrock open models updated in this PR for the same Converse behavior. Action: Add interleaved = true (or the correct field form) for qwen.qwen3-32b-v1:0, or explain why reasoningContent is not a usable interleaved channel here.
• [medium] [possible mistake] providers/amazon-bedrock/models/qwen.qwen3-vl-235b-a22b.toml:5 - Check: Host structured_output overrides must reflect verified API capability. Why: Lab metadata has structured_output = true, and this PR flips the Bedrock entry to false with only model-card/pricing comments—no live or docs note that structured/JSON output fails on this route. Action: Cite host evidence that structured output is unsupported on Bedrock for this model, or drop the override and inherit the lab value.
• [low] [possible mistake] providers/amazon-bedrock/models/minimax.minimax-m2.1.toml:8 - Check: Material capability flips should be evidenced. Why: All three MiniMax entries change structured_output from false to true, but first-party MiniMax entries do not claim it and the leading comments only cover thinking/pricing/limits—not structured output. Action: Add a short leading note of the Bedrock check that confirms structured output, or leave the field unset/false until verified.

[rekram1-node]

Addressed the second review pass in 7b23de6 with direct Converse verification:

• Added interleaved = true to MiniMax M2, M2.1, and M2.5. All three returned reasoningContent without a caller reasoning setting.
• Added interleaved = true to Qwen 3 32B, matching the existing live high/low evidence.
• Confirmed structured output on all three MiniMax models: a contradictory prose instruction plus outputConfig.textFormat produced valid schema-constrained JSON.
• Confirmed the same structured-output behavior on Qwen 3 VL 235B, so removed its incorrect structured_output = false host override and now inherit the lab’s true capability.
• Added concise live-evidence notes to the affected headers.

Sanitized results: MiniMax M2/M2.1/M2.5 and Qwen VL all returned HTTP 200; MiniMax responses included reasoning blocks and all four returned schema-valid {ok: boolean} JSON.

bun validate and git diff --check pass.

[rekram1-node]

Superseded by three smaller PRs based on current dev after the broad reviewer repeatedly hung:

• fix(amazon-bedrock): correct Gemma catalog #6824 — Google/Gemma (9 files)
• fix(amazon-bedrock): correct DeepSeek and Qwen catalog #6823 — DeepSeek/Qwen (11 files)
• fix(amazon-bedrock): correct hosted reasoner catalog #6822 — xAI/NVIDIA/MiniMax/Z.AI/Moonshot (16 files)

All findings from both completed review passes are included, and each split passes bun validate independently.