nexify 版本演进报告 · 基础 v11 → v21

基础版本 v11 → 当前 v21 · 语料:codewhale / zed / pdf · 生成 2026-09-06T02:01 UTC

口径注:v14 起 UNRESOLVED 不再计入 Rust std / Python builtin 等合理不绑类(指标治理 post-11 · bb8989e),故 v13→v14 的 UNRESOLVED 断崖与绑定率跳涨属口径变化,非召回提升;INFERRED resolved / AMBIGUOUS / 哨兵逐位零变化是真实约束。

1. codewhale(Rust workspace 主回归基准)

指标v11v12v13v14v15v16v17v18v21Δ(最后两版)
EXTRACTED / extracted40,67340,67340,67340,67340,67340,85240,70740,70740,669−38
推断已解析(INFERRED resolved)21,62222,45122,45122,45122,53422,48122,65823,51223,496−16
推断未决(INFERRED extracted)6,6216,6216,6216,6216,6216,6216,6216,6216,6210
歧义(AMBIGUOUS)25225225225225325325313299−33
未解析(UNRESOLVED)33,95333,12433,1244,9014,8174,7084,4894,4893,547−942
绑定率(仅 Rust 推断面)38.7%40.2%40.2%81.3%81.6%81.9%82.7%83.6%86.6%+3.0%
冷建图耗时中位数(s)—33.36s32.71s33.35s32.55s32.71s26.32s26.94s29.63s+2.69s
解析耗时(s)—7.59s6.92s6.81s7.17s7.89s4.15s4.16s4.55s+0.39s
峰值内存(MB)——————1151151150

2. zed(Rust 超大 workspace)

指标v11v12v13v14v15v16v17v18v21Δ(最后两版)
EXTRACTED / extracted30,27630,27630,27630,27630,27631,08430,43530,43530,4350
推断已解析(INFERRED resolved)31,65136,15137,21837,21839,83839,78539,68981,94481,9440
推断未决(INFERRED extracted)13,68113,68113,68113,68113,68113,68113,68113,68113,6810
歧义(AMBIGUOUS)8,1048,1046,7526,7526,7526,7526,7522,0742,0740
未解析(UNRESOLVED)43,39838,89838,8986,1933,6222,9002,5802,5802,5800
绑定率(仅 Rust 推断面)38.1%43.5%44.9%74.2%79.3%80.5%81.0%94.6%94.6%0
冷建图耗时中位数(s)—79.64s80.62s76.15s75.94s77.73s50.3s50.99s58.36s+7.37s
解析耗时(s)—34.41s34.27s32.49s31.92s33.14s13.35s13.22s23.82s+10.6s
峰值内存(MB)——————1331331330

3. pdf(Rust 单 package)

指标v11v12v13v14v15v16v17v18v21Δ(最后两版)
EXTRACTED / extracted3,1303,1303,1303,1303,1303,1433,0933,0933,0930
推断已解析(INFERRED resolved)6036036036036036038549169160
推断未决(INFERRED extracted)1531531531531531531531531530
歧义(AMBIGUOUS)4444447660
未解析(UNRESOLVED)1,5531,5531,5535515515383043043040
绑定率(仅 Rust 推断面)27.9%27.9%27.9%52.1%52.1%52.7%73.3%74.7%74.7%0
冷建图耗时中位数(s)—2.13s1.4s1.47s1.46s1.5s1.48s1.63s1.64s+0.01s
解析耗时(s)—0.26s0.18s0.19s0.22s0.19s0.18s0.19s0.17s−0.03s
峰值内存(MB)——————3131310

4. 召回哨兵(codewhale,不得回退)

哨兵v11v12v13v14v15v16v17v18v21判定
ToolContext::new 构造器617617617617617617617617617保持
ViewStack::push 方法232323232323232323保持
user_home 跨 crate161616161616161616保持
FleetRunId::from 限定路径575757575757577676保持
re-export 五符号主哨兵818181818181818181保持
RuntimeToolServices::default999999999保持
drain_pending_steers 方法666666666保持
workspace 子模块直导272727272727272727保持
guard 内调用111111111保持
discover 跨模块555555555保持

色义:绿 = 进化(耗时下降 / 召回上升 / 漏绑下降 / 哨兵保持);红 = 退化(耗时上升 / 哨兵回退);灰 = 持平或缺失。耗时来自各版本 scan 采样中位数(≥2 次);Δ=0 灰,无噪声特判。

5. UNRESOLVED 归因(v21 当前版快照)

codewhale 真漏绑 top:lock_test_env 404 · path 150 · post 65 | 外部依赖 top:tempdir 1154 · method 169 · Json 141

zed 真漏绑 top:rust_lang 89 · Unclipped 31 · neutral 24 | 外部依赖 top:assertEqual 112 · add_argument 103 · join_all 52

pdf 真漏绑 top:extract_pages_markdown 2 · extract_text 1 | 外部依赖 top:assertEqual 21 · add_argument 20 · write_text 16

语料真漏绑跨语言桥接外部依赖合计
codewhale653(18.4%)132(3.7%)2762(77.9%)3547
zed585(22.7%)304(11.8%)1691(65.5%)2580
pdf3(1.0%)75(24.7%)226(74.3%)304

术语:真漏绑 = 仓库内同语言定义存在但未绑(该打);跨语言桥接 = 目标定义在仓库内但属另一语言(缺桥接识别,该绑);外部依赖 = 仓库内无定义(第三方库/内置/生态——解析自包含外,合理不绑)。当前版快照不进演进列。

6. 候选改善清单(按推荐价值排序)

#内容规模推荐价值验收指针
长期references 类型使用边 + 社区/枢纽榜(Leiden/god-nodes)—待评估M5 候选
长期≤2 字符裸标签白名单(scoped use 已解决,剩无 use 裸标签)825 条基线待评估常见名保护残余短板
长期TS receiver 类型证明(成员边恢复评估通道)原成员池 ~60%undefinedTODO-M5 §6.2;前提:import 链(M5-3)落地后 receiver=导入符号/显式注解/new 构造三类可证子集量化

循环:本报告 → 从中选取推荐价值最高的候选派活 → 子代理开发 → 新版本 v22 → 语料全量扫描 + 归档 → 生成 nexify-v11-v22-report.html(数据叠加)。推荐价值 = 收益(召回/精度/规模)÷ 成本(改动量/风险)。