fix(claude): handle max-token Responses truncation

[Reekin]

Summary

• default Fable Responses requests to 64K output tokens while respecting lower registered model limits
• preserve explicit max_output_tokens and document model-scoped payload.override configuration for budgets such as 128K
• map Claude stop_reason: max_tokens to response.incomplete with incomplete_details.reason: max_output_tokens
• preserve partial reasoning/text and keep already completed message item statuses stable
• apply the terminal mapping to both streaming and non-streaming Responses paths

Problem

Claude can consume the full output budget in reasoning without producing a visible assistant message. The Claude-to-Responses translator currently ignores message_delta.delta.stop_reason and always emits response.completed, so Codex can silently finalize the turn with no final answer.

Configuration

payload:
  override:
    - models:
        - name: "claude-fable-5-1"
          protocol: "claude"
      params:
        max_tokens: 128000

The override is applied after protocol translation and works for both OAuth and API-key Claude execution.

Verification

• go test ./internal/translator/claude/openai/responses -count=1
• go test ./internal/runtime/executor -run TestClaudeExecutorPayloadOverrideMaxTokens -count=1
• go build -o <temp>/cli-proxy-api ./cmd/server
• real HTTP smoke: /v1/responses -> configured Claude upstream received max_tokens: 128000; downstream ended with response.incomplete and preserved reasoning output

go test ./... -count=1 was also exercised. Changed packages pass; the existing XAI TTFT timing assertion can intermittently report ttft = 0s on Windows and passes independently.