fix(chat): persist assistant reasoning_content across turns

[cursor]

Summary

• DeepSeek thinking-mode APIs return HTTP 400 when a later request includes tools but prior assistant messages omit reasoning_content.
• SkillLite already kept the field in memory during one agent turn, then dropped it on transcript persist/reload. The next skilllite chat / desktop turn in the same session rebuilt history without the field.
• Fix: store optional reasoning_content on transcript assistant Message rows, restore it in transcript_entry_to_message, and keep outbound OpenAI JSON echoing it.

Trigger: configure a DeepSeek thinking model, run skilllite chat --session s --message "..." (tools are sent), then a second turn on the same session. Before the fix the second request omitted reasoning_content and the API rejected it.

session_key path traversal was also reported in this sweep; that issue is already covered by open PR #126 and was not re-opened.

Task Linkage

• Task ID: TASK-2026-071
• Task folder: tasks/TASK-2026-071-persist-reasoning-content/

Injected Specs

• spec/verification-integrity.md
• spec/task-artifact-language.md
• spec/architecture-boundaries.md (reviewed; no dependency/layering change)
• spec/security-nonnegotiables.md (not a sandbox/security policy change)
• spec/testing-policy.md
• spec/docs-sync.md (N/A — crash/API-contract fix, no command/env/docs change)
• spec/rust-conventions.md
• spec/structured-signal-first.md
• spec/capability-gap-evolution.md

Validation Evidence

• Commands executed:
  • cargo test -p skilllite-executor --lib reasoning — 2 passed
  • cargo test -p skilllite-agent reasoning — reload + OpenAI echo tests passed
  • cargo test -p skilllite-agent — 250 passed; 0 failed
  • cargo fmt --check
  • cargo clippy --all-targets -- -D warnings
  • cargo test — workspace crates ok
  • python3 scripts/validate_tasks.py — Task validation passed (71 task directories checked).
• Key results:
  • Legacy transcript rows without the field still deserialize as None.
  • Reloaded assistant messages include CJK reasoning_content.
  • openai_api_message echoes the field on assistant payloads.

Regression Scope

• Areas likely affected:
  • Chat transcript persist/reload and DeepSeek thinking multi-turn follow-ups.
• Explicit non-goals:
  • Session-key path validation (PR fix(chat): reject path-escaping session keys #126).
  • Persisting intermediate tool-call assistant rows.
  • Desktop thinking-text UI.

Docs Sync (EN/ZH)

• Not needed
• Updated EN + ZH docs
• Files:
  • N/A

Review Checklist

• Acceptance criteria in tasks/TASK-2026-071-persist-reasoning-content/TASK.md satisfied
• tasks/TASK-2026-071-persist-reasoning-content/STATUS.md updated with latest progress
• tasks/TASK-2026-071-persist-reasoning-content/REVIEW.md includes merge readiness decision
• tasks/board.md status is up to date