[projection-agent] Publish NFugu JudgeEval codex fuse follow-up

docs/nfugu-judge-eval-v0.md

@@ -1,14 +1,19 @@
 # NFugu JudgeEval v0
 
-> **Status: Paused (2026-07-02).** Keep this harness as a v0 prototype. Do
-> not run new paid/remote batches until it supports single-baseline injection,
-> pool-wise candidate judging, executable code verification for code tasks, and
-> explicit selection-regret metrics.
+> **Status: NFugu paused; CodexFuse sanity-check enabled (2026-07-02).**
+> Keep the NFugu harness as a v0 prototype. Do not run new paid/remote NFugu
+> batches until it supports single-baseline injection, pool-wise candidate
+> judging, executable code verification for code tasks, and explicit
+> selection-regret metrics. `codex_fuse` is allowed as a smaller monotonicity
+> sanity-check because it reuses the exact `single` answer as C0, generates a
+> second Codex answer C1, asks Codex to synthesize F1 from the strengths of C0
+> and C1, then conservatively selects from C0/C1/F1.
 
 NFugu JudgeEval v0 is a fast, judge-based eval harness for comparing:
 
 ```text
 single_codex
+codex_fuse
 best_of_2n_remote
 n_plus_n_remote
 ```
@@ -37,6 +42,7 @@ custom manifest can be passed with `--manifest`.
 
 ```text
 single      codex-cli
+fuse        codex-cli second candidate + codex-cli fusion + codex-cli conservative pool judge
 workers     codex-cli,ark-plan-deepseek
 optimizers  agy-cli,ark-plan-glm
 judge       ark-plan-minimax
@@ -49,8 +55,9 @@ kept outside the candidate pool by default.
 
 For each task, the harness:
 
-1. runs `single`, `best_of_2n`, and `n_plus_n` through `scripts/run_nfugu_task.py`;
-2. anonymizes the three final answers as `A`, `B`, and `C`;
+1. runs the requested condition set; the default remains `single`,
+   `best_of_2n`, and `n_plus_n`;
+2. anonymizes the condition final answers as `A`, `B`, `C`, ...;
 3. asks the judge for `winner_label`, `runner_up_label`, scores, failure reasons,
    confidence, and rationale;
 4. runs secondary static verification for code tasks;
@@ -59,6 +66,31 @@ For each task, the harness:
 The judge never sees condition names. The label map is stored only in the local
 result artifact.
 
+`codex_fuse` is special: it is not another independent full run. It uses the
+already generated `single` final answer as C0, generates one additional
+`codex-cli` answer as C1, asks a fusion Codex call to produce F1 by combining
+the real strengths of C0 and C1, then selects from C0/C1/F1 with verifier-first
+checks and a conservative blind pool judge. If the judge does not strongly
+select C1 or F1, the final answer stays C0.
+
+`codex_double` follows the same audit principle without the fusion call: C0 is
+the exact `single` answer and C1 is one additional Codex answer. When the
+condition set is `single,codex_double`, the report-level `winner` is C0 or C1,
+not the synthetic `codex_double` condition name. If the monotonic guard keeps
+C0, the row records that the final answer equals the single baseline instead of
+claiming a separate condition-level win.
+
+When the condition set is `single,codex_fuse`, the report-level `winner` is the
+best candidate from the C0/C1/F1 pool, not the condition name. This avoids
+misreading "codex_fuse won" as "F1 won" when the actual selected candidate was
+C1.
+
+Rows in `results.jsonl` carry a run signature covering the manifest hash,
+condition order, provider preset/provider slots, judge providers, thresholds,
+and generation limits. Cached rows are reused only when that signature matches,
+so changing from NFugu conditions to `codex_fuse` or `codex_double` cannot
+silently mix stale rows into the active summary/report.
+
 ## Commands
 
 Prepare the default manifest without model calls:
@@ -79,6 +111,20 @@ python -B scripts/run_nfugu_judge_eval.py `
   --out-dir .memsuos/evals/nfugu-judge-eval-v0-smoke
 ```
 
+Run the CodexFuse sanity-check shard:
+
+```powershell
+python -B scripts/run_nfugu_judge_eval.py `
+  --conditions single,code_fuse `
+  --limit 2 `
+  --single-provider codex-cli `
+  --second-provider codex-cli `
+  --fuse-provider codex-cli `
+  --fuse-judge-provider codex-cli `
+  --judge-provider codex-cli `
+  --out-dir .memsuos/evals/codex-fuse-judgeeval-v0
+```
+
 Run the real first shard:
 
 ```powershell

docs/project-spec-index.md

@@ -159,7 +159,7 @@ memsuOS 当前聚焦的是：
 | `docs/NFuguMode.md` | NFuguMode 设计（已暂停实现） | N+N 候选优化赛：V1 runner 与 JudgeEval v0 保留为研究/原型；2026-07-02 暂停继续实现和远端付费评测，重启前需补 baseline 注入、pool-wise judge/verifier 和 selection-regret 报告 |
 | `docs/fugu-first-eval-plan.md` | FuguMax 首批评测计划（条件已弃用） | Aider/BFCL 轻量阶段的评测准备和失败实现复盘；抽样与打分契约沿用到 NFuguMode 首批评测（N+N vs best-of-2N 消融） |
 | `docs/nfugu-first-eval-runbook.md` | NFugu 首批评测 runbook（暂停） | 首批 eval 操作记录；暂停从该 runbook 直接开跑，等待 baseline 注入、pool-wise candidate scoring 和 selection-regret 指标补齐 |
-| `docs/nfugu-judge-eval-v0.md` | NFugu JudgeEval v0（暂停） | 20 题三方 blind judge 原型；默认 agy-cli 替代 kimi-cli，但暂停新远端批次，等待单调性保护和代码题可执行验证补齐 |
+| `docs/nfugu-judge-eval-v0.md` | NFugu JudgeEval v0（NFugu 暂停，CodexFuse sanity-check 可用） | 20 题三方 blind judge 原型；默认 agy-cli 替代 kimi-cli；NFugu 远端批次暂停，新增 `codex_fuse` 条件用于 C0/C1/F1 同构 Codex 融合 sanity-check |
 
 ## 7. 源码索引