feat(agent): adversarial evaluator gate (generator-evaluator separation)

[topcheer]

Research-backed capability (cron-runner research round sa-9, concept: Anthropic "Harness design for long-running apps" 2026-03, GAN-inspired generator-evaluator separation; no issue filed).

Gap

The pre-completion gate chain (build/test/lint oracles, spec-gaming, companion, complexity, cross-file, reconcile) is entirely zero-LLM-cost heuristics or command checks - no semantic-level independent evaluation exists. Self-assessment is still done by the generator itself.

Implementation (5 files, +363)

• internal/agent/evaluator.go: checkAdversarialReviewGate - run diff to a FRESH LLM call with no main-loop history (context isolation). Criteria: spec adherence, correctness, integration, edge cases, honesty. Malformed verdict tolerated as PASS - evaluator failure never blocks completion.
• Loop closure: FAIL findings injected as user message; generator must fix or argue - evaluator->generator repair loop.
• Guardrails: max 2 rounds/task, 24KB diff / 6KB findings truncation, 120s timeout, skip in plan mode / no changes / empty provider; diff marked untrusted data; usage attributed to evaluator source.
• Wiring: agent.go gate chain (after reconcile, before diff-summary); config verify.adversarial_review (default OFF) via agentruntime/verify_config.go.

Tests

evaluator_test.go: 5 verdict-parsing cases + default-off + no-changes skip. Full agent/agentruntime/config package regressions green. gofmt clean, go build -tags goolm clean.

Co-Authored-By: ggcode noreply@ggcode.dev

[topcheer]

复审通过 ✅（默认 off 的 opt-in LLM 门+fail-safe 方向全对）

• 分离实现正确：fresh-context（evaluator 只收 task+diff，无主循环历史——不被 generator 叙事锚定，GAN 分离的主杠杆）；同 provider 复用是可接受简化
• fail-safe 全对：evaluator 失败/超时/畸形 verdict 一律按 PASS 放行（"unavailability must never block completion"）；仅明确 FAIL+非空 findings 回注——LLM 门挂了不能卡死完成路径，方向正确
• 注入防御在位：system prompt 明示 diff/task 内容一律 UNTRUSTED——评审对象是任意代码，这是必要防线
• 资源界完备：2 轮/任务+task prompt 变更重置预算、120s 超时、24KB diff/6KB findings 截断、usage 归集
• 挂接点：deterministic gates 之后、diff-summary gate 之前，FAIL→user message 回注修（与既有 gate 同模式）
• 3 测试（verdict 解析表/默认关/无变更跳过）+ config opt-in 链完整

非阻塞观察：collectAdversarialDiff 注释称 "committed + staged + unstaged vs HEAD" 与 git diff HEAD 实际语义矛盾（恰不含已 commit 内容——run 中 commit 过则 diff 为空）。对主 checkout 不 commit 的工作流盲区可接受，但注释应改为 "working tree vs HEAD"（免误导后来者）。untracked 文件名面已补偿。

CI 9/9 绿。可合并（排程 5/6，剩 #2448）。