HermitCrab analysis: dead-internal work is material only on Aweti (39.8%) and v1 (12.2%); shadow-mode summary not run (research lead)

未关闭
  • 哪里不对？
#605 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
48/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
活跃
技术栈
csharp
领域
performance

调研方向

Read the C1 discussion in docs/hermitcrab-2026-10-21-search-cost-round.md (lines 114–168) and lead 2 (lines 319–330), then trace the C# summary and work-step measurement entry points. Run the summary in shadow mode without changing parse results; done means reporting how much plain dead-internal bucket work it would decide and the summary's cost.

由索引模型根据 Issue 内容生成。

描述

Claude-drafted.

The plain dead-internal bucket, the only analysis work a sound summary of reachable roots could decide at a profit, is material on two of eight lists: 39.8% of parser work on Aweti and 12.2% on a private v1 held list, and 0-5% on the other six. A shadow-mode run has not been done.

Status: research lead (C1), ceiling measured, not built. It reopens only through the experiment below.

What we know (HC-Rust port, hindsight census, parser work in work steps):

  • The plain bucket is 0-5% on six of eight lists. Aweti's 39.8% is on 183 completed words of 206; v1 is 12.2% of 105.8M work steps over 7,206 words.
  • Most dead work cannot be pruned: it sits in dead leaves, or in nodes that could still split off a compound non-head. Amharic, Sena and Mbugwe put 36-46% of parser work there.
  • Any sound summary must model reduplication, infixes, truncation, deletion re-insertion and compounding before it may prune.
  • The census builds predate change C, so template-heavy shares use an older, larger denominator.

Not known: the summary's own cost, and the share of plain-bucket work it would decide.

Proposed change: run the summary in shadow mode and count, in work steps, how much bucket work it would decide and what it costs. Parse results do not change. Close the C1 research lead if the decided share is small.

Related: #485 (open), #562 (open). Retired record: PR #490 (closed). PanGloss: sillsdev/PanGloss#107 (pointer, open).

Next: build the shadow summary on Aweti and v1 and report decided share and own cost in work steps.

Evidence -- dead-work census by list

A hindsight census charged each unapplication's steps and work to its input node; a node is live when a lookup on it, on an alias or on a descendant found a root. Share of parser work:

List Parser work Lexically dead Dead leaves Dead, may compound Dead internal, plain
Private base, held 50,345,810 22.6% 6.1% 14.6% 1.9%
Private v1, held 105,768,202 38.7% 19.9% 6.6% 12.2%
Private deletion variant, held, before fix B 303,966,311 16.0% 12.7% 0.0% 3.4%
Indonesian 1,906,012 41.0% 15.3% 23.8% 1.9%
Amharic 39,395,506 66.1% 20.3% 45.8% 0.0%
Sena 271,894,826 55.5% 11.9% 39.0% 4.7%
Mbugwe 610,569,774 45.5% 9.7% 35.8% 0.0%
Aweti 488,089,456 74.9% 35.1% 0.0% 39.8%

Amharic: 669 words completed. Sena: 5,134 completed at a 30 s wall limit. Mbugwe: first 400, 394 completed. Aweti: 183 of 206 completed at the same limit. Dead leaves cost about as much to prove dead as to run; a may-compound node is the disqualifier of ledger row 10.

Sources -- pinned to 2eba2f50 (closed PR #490 record)
What would reopen this

A shadow run where a sound summary decides a meaningful share of the plain bucket on Aweti or v1, at a summary cost that does not cancel the saving.

主要语言
C#
星标
29
派生
16
平均合并
6 天 7 小时
30 天内合并 PR
14

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 有 Pull Request 模板
  • 没有贡献指南

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

sillsdev/machine 的其他 Issue

查看 sillsdev/machine 的全部 Issue

相似的 Issue

更多 C# Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。