docs: survey code agent benchmarks and their headless interface requirements

[minixalpha]

Two research notes on code-agent benchmarks — bilingual, hand-written Chinese
source plus generated English — and the 0.8.x changelog they open.

What's here

code_agent_benchmark.md — a survey of the code-agent benchmarks reported
across six 2026 model releases (DeepSeek-V4-Flash-0731, Claude Opus 5, GPT-5.6
Sol, Qwen3.8-27B, Kimi-K3, GLM-5.2), flattened into one table. Terminal-Bench
2.1 and DeepSWE turn out to be the only two every release has a number for. It
closes with a recommendation for this project — Terminal-Bench 2.1 first, a
SWE-bench Verified subset second, NL2Repo-Bench third — and a P0/P1/P2 list of
what nanoPyCodeAgent still lacks.

benchmark_headless_interface.md — a follow-up that reads the integration
surface of those three benchmarks down to the source level:

• Terminal-Bench 2.1 / Harbor: the BaseInstalledAgent adapter interface,
  both official invocation examples (Claude Code pipes the instruction over
  stdin, mini-swe-agent passes --task=), and eight constraints that are in
  the source but not in the documentation.
• SWE-bench Verified: no agent interface at all — only the predictions
  format. The runner-side essentials (image naming, /testbed, step/cost
  budgets, patch collection) come from mini-swe-agent.
• NL2Repo-Bench: scoring is decoupled from the agent, so the only requirement
  is running one prompt headlessly in a given directory; integrating means
  forking its OpenHands runner.

The most important finding: a harness reads a non-zero exit code as its own
failure, so an agent that runs out of turns without solving the task must
still exit 0. The first draft of the survey had this backwards.

Revisions to the earlier survey

• Exit-code semantics inverted (P0-2).
• The working-directory item moved from P0 to P2 — all three benchmarks deliver
  it through the container's default WORKDIR — with P1/P2 renumbered.
• The retry requirement simplified: Harbor classifies API errors itself and
  retries via --retry-include, so the agent only needs basic backoff plus
  printing errors verbatim.
• The DeepSWE leaderboard wording corrected: it pins the harness to
  mini-swe-agent for consistency rather than gating submissions on it, which
  means a self-built harness has no place on that leaderboard.
• Terminal-Bench's reasons reordered — the strongest one is that its
  leaderboard entries are two-dimensional (harness + model), so this project
  has a legitimate place on it.

Docs bookkeeping

• New docs/changelogs/0.8.x.md, with both surveys under [Unreleased].
• Both notes indexed in docs/research/README.md.
• The English version of each note is in sync with its Chinese source.

No code changes.