fix(runtime): honor selected model output limits

[Astro-Han]

Summary

• Resolve the selected model's output-token limit from the connection override or the canonical model metadata.
• Pass that limit to every AI SDK streamText request.
• Cover Kimi K3's 131,072-token contract through the AiSdkBackend.send() boundary.

Verification

• RED: the focused regression observed maxOutputTokens === undefined on current main instead of 131072.
• npm --workspace @maka/runtime test — 2,112 passed, 7 skipped, 0 failed.
• git diff --check

The repository-wide bootstrap rebuild was not used as a gate because this checkout's shared dependency installation is missing @modelcontextprotocol/sdk, which stops the unrelated @maka/mcp build. The complete runtime workspace build and test suite passed.

Root cause

The runtime selected the K3 model correctly but did not forward its metadata-backed output limit to streamText. The AI SDK/provider adapter therefore fell back to its unknown-model behavior, allowing long responses to terminate at the wrong max_tokens boundary.

Scope

This PR changes only the runtime model request contract. Harbor task deadlines, harness recovery, provider telemetry, and benchmark reporting remain separate follow-ups.

[Astro-Han]

Summary

• Resolve the selected model's output-token limit from the connection override or the canonical model metadata.
• Pass that limit to every AI SDK streamText request.
• Cover Kimi K3's 131,072-token contract through the AiSdkBackend.send() boundary.

Verification

• RED: the focused regression observed maxOutputTokens === undefined on current main instead of 131072.
• npm --workspace @maka/runtime test — 2,112 passed, 7 skipped, 0 failed.
• git diff --check

The repository-wide bootstrap rebuild was not used as a gate because this checkout's shared dependency installation is missing @modelcontextprotocol/sdk, which stops the unrelated @maka/mcp build. The complete runtime workspace build and test suite passed.

Root cause

The runtime selected the K3 model correctly but did not forward its metadata-backed output limit to streamText. The AI SDK/provider adapter therefore fell back to its unknown-model behavior, allowing long responses to terminate at the wrong max_tokens boundary.

Scope

This PR changes only the runtime model request contract. Harbor task deadlines, harness recovery, provider telemetry, and benchmark reporting remain separate follow-ups.

[Astro-Han]

Summary

• Resolve the selected model's output-token limit from the connection override or the canonical model metadata.
• Pass that limit to every AI SDK streamText request.
• Cover Kimi K3's 131,072-token contract through the AiSdkBackend.send() boundary.

Verification

• RED: the focused regression observed maxOutputTokens === undefined on current main instead of 131072.
• npm --workspace @maka/runtime test — 2,112 passed, 7 skipped, 0 failed.
• git diff --check

The repository-wide bootstrap rebuild was not used as a gate because this checkout's shared dependency installation is missing @modelcontextprotocol/sdk, which stops the unrelated @maka/mcp build. The complete runtime workspace build and test suite passed.

Root cause

The runtime selected the K3 model correctly but did not forward its metadata-backed output limit to streamText. The AI SDK/provider adapter therefore fell back to its unknown-model behavior, allowing long responses to terminate at the wrong max_tokens boundary.

Scope

This PR changes only the runtime model request contract. Harbor task deadlines, harness recovery, provider telemetry, and benchmark reporting remain separate follow-ups.

[Astro-Han]

Summary

• Resolve the selected model's output-token limit from the connection override or the canonical model metadata.
• Pass that limit to every AI SDK streamText request.
• Cover Kimi K3's 131,072-token contract through the AiSdkBackend.send() boundary.

Verification

• RED: the focused regression observed maxOutputTokens === undefined on current main instead of 131072.
• npm --workspace @maka/runtime test — 2,112 passed, 7 skipped, 0 failed.
• git diff --check

The repository-wide bootstrap rebuild was not used as a gate because this checkout's shared dependency installation is missing @modelcontextprotocol/sdk, which stops the unrelated @maka/mcp build. The complete runtime workspace build and test suite passed.

Root cause

The runtime selected the K3 model correctly but did not forward its metadata-backed output limit to streamText. The AI SDK/provider adapter therefore fell back to its unknown-model behavior, allowing long responses to terminate at the wrong max_tokens boundary.

Scope

This PR changes only the runtime model request contract. Harbor task deadlines, harness recovery, provider telemetry, and benchmark reporting remain separate follow-ups.

[Astro-Han]

Summary

• Resolve the selected model's output-token limit from the connection override or the canonical model metadata.
• Pass that limit to every AI SDK streamText request.
• Cover Kimi K3's 131,072-token contract through the AiSdkBackend.send() boundary.

Verification

• RED: the focused regression observed maxOutputTokens === undefined on current main instead of 131072.
• npm --workspace @maka/runtime test — 2,112 passed, 7 skipped, 0 failed.
• git diff --check

The repository-wide bootstrap rebuild was not used as a gate because this checkout's shared dependency installation is missing @modelcontextprotocol/sdk, which stops the unrelated @maka/mcp build. The complete runtime workspace build and test suite passed.

Root cause

The runtime selected the K3 model correctly but did not forward its metadata-backed output limit to streamText. The AI SDK/provider adapter therefore fell back to its unknown-model behavior, allowing long responses to terminate at the wrong max_tokens boundary.

Scope

This PR changes only the runtime model request contract. Harbor task deadlines, harness recovery, provider telemetry, and benchmark reporting remain separate follow-ups.

[Astro-Han]

Summary

• Resolve the selected model's output-token limit from the connection override or the canonical model metadata.
• Pass that limit to every AI SDK streamText request.
• Cover Kimi K3's 131,072-token contract through the AiSdkBackend.send() boundary.

Verification

• RED: the focused regression observed maxOutputTokens === undefined on current main instead of 131072.
• npm --workspace @maka/runtime test — 2,112 passed, 7 skipped, 0 failed.
• git diff --check

The repository-wide bootstrap rebuild was not used as a gate because this checkout's shared dependency installation is missing @modelcontextprotocol/sdk, which stops the unrelated @maka/mcp build. The complete runtime workspace build and test suite passed.

Root cause

The runtime selected the K3 model correctly but did not forward its metadata-backed output limit to streamText. The AI SDK/provider adapter therefore fell back to its unknown-model behavior, allowing long responses to terminate at the wrong max_tokens boundary.

Scope

This PR changes only the runtime model request contract. Harbor task deadlines, harness recovery, provider telemetry, and benchmark reporting remain separate follow-ups.

[Astro-Han]

Summary

• Resolve the selected model's output-token limit from the connection override or the canonical model metadata.
• Pass that limit to every AI SDK streamText request.
• Cover Kimi K3's 131,072-token contract through the AiSdkBackend.send() boundary.

Verification

• RED: the focused regression observed maxOutputTokens === undefined on current main instead of 131072.
• npm --workspace @maka/runtime test — 2,112 passed, 7 skipped, 0 failed.
• git diff --check

The repository-wide bootstrap rebuild was not used as a gate because this checkout's shared dependency installation is missing @modelcontextprotocol/sdk, which stops the unrelated @maka/mcp build. The complete runtime workspace build and test suite passed.

Root cause

The runtime selected the K3 model correctly but did not forward its metadata-backed output limit to streamText. The AI SDK/provider adapter therefore fell back to its unknown-model behavior, allowing long responses to terminate at the wrong max_tokens boundary.

Scope

This PR changes only the runtime model request contract. Harbor task deadlines, harness recovery, provider telemetry, and benchmark reporting remain separate follow-ups.

[Astro-Han]

Summary

• Resolve the selected model's output-token limit from the connection override or the canonical model metadata.
• Pass that limit to every AI SDK streamText request.
• Cover Kimi K3's 131,072-token contract through the AiSdkBackend.send() boundary.

Verification

• RED: the focused regression observed maxOutputTokens === undefined on current main instead of 131072.
• npm --workspace @maka/runtime test — 2,112 passed, 7 skipped, 0 failed.
• git diff --check

The repository-wide bootstrap rebuild was not used as a gate because this checkout's shared dependency installation is missing @modelcontextprotocol/sdk, which stops the unrelated @maka/mcp build. The complete runtime workspace build and test suite passed.

Root cause

The runtime selected the K3 model correctly but did not forward its metadata-backed output limit to streamText. The AI SDK/provider adapter therefore fell back to its unknown-model behavior, allowing long responses to terminate at the wrong max_tokens boundary.

Scope

This PR changes only the runtime model request contract. Harbor task deadlines, harness recovery, provider telemetry, and benchmark reporting remain separate follow-ups.