fix(reasoning): honor configured summary mode

[duclvz]

Summary

• Honor CODEX_CONFIG.model_reasoning_summary when starting prompt turns.
• Support auto, concise, detailed, and none.
• Preserve forced none for API-key authentication and models without reasoning support.
• Keep auto as the fallback so thinking chunks remain visible when no mode is configured.
• Document the configuration in README.md and readme-dev.md.

Problem

CodexAcpClient.sendPrompt() always sent summary: "auto" for supported models. Because turn/start.summary is a per-turn override, this replaced an explicit reasoning-summary mode supplied through CODEX_CONFIG.

Removing the field entirely also caused the default thinking block to disappear because no reasoning-summary chunks were emitted in the unconfigured production path.

Solution

Resolve the turn summary mode using this precedence:

1. Compatibility guard → none
2. Valid CODEX_CONFIG.model_reasoning_summary
3. Default → auto

This preserves existing behavior while allowing users to select a more specific summary mode.

Scope

This change reads model_reasoning_summary from CODEX_CONFIG. It does not add separate resolution of ~/.codex/config.toml.

Testing

• TypeScript typecheck passed
• Build passed
• 303 tests passed; 28 skipped
• Live auto and detailed runs both emitted ACP thought chunks and completed successfully

Refs #294