docs: record the deepseek-v4.1-flash rerun of six failed tasks

[minixalpha]

What

Record a benchmark experiment: re-run the six tasks that scored reward 0 in
tb21-stream-recovery-pilot20-20260921, using deepseek/deepseek-v4.1-flash
with every other limit unchanged (agent 1386df48e, max_tokens=65536,
max_turns=50, 3600 s per task, concurrency 2, one attempt, no retries).

Result

• 4 / 6 passed: kv-store-grpc, dna-assembly, log-summary-date-ranges,
  circuit-fibsqrt.
• 2 / 6 still failed by 50-turn exhaustion: caffe-cifar-10 (training only
  reached Iteration 200), mteb-leaderboard (never wrote /app/result.txt).
• dna-assembly and circuit-fibsqrt passed cleanly for the first time.
• 206 / 206 model replies returned complete; no transport interruptions;
  total cost $0.278759041.

Changes

• docs/dev_notes/zh-CN/0.8.x.md — new “更换模型复跑未通过的 6 题” section (Chinese source).
• docs/dev_notes/en/0.8.x.md — English version generated from the Chinese source.
• benchmarks/harbor/results/tb21-v41flash-6task-20260925.json — structured result with per-task outcome, tokens, and cost.

Notes

• No changelog entry: this records a benchmark experiment, not a user-facing change.
• Not a controlled comparison — the model changed while routing, sampling, cache,
  and backend stayed uncontrolled, with a single attempt per task.
• Raw trials and the experiment record live in Git-ignored jobs/ and are not uploaded.