Skip to content

m25-plan-typeduck-web-dogfooding-round-2.md

Latest commit

 

History

History
594 lines (365 loc) · 65.2 KB

m25-plan-typeduck-web-dogfooding-round-2.md

File metadata and controls

594 lines (365 loc) · 65.2 KB
Raw

M25 TypeDuck-Web Dogfooding Round 2 Implementation Plan

Status: Complete - Milestone: M25 (TypeDuck-Web dogfooding round 2) - Created: 2026-06-21 - Closed: 2026-06-21 - Type: archived issue ledger / execution record

For agentic workers: REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (- [ ]) syntax for tracking.

Intake note: This plan starts the second manual dogfooding loop for the internal TypeDuck-Web playground. Add the user's upcoming feedback as M25-DOGFOOD-* rows here before implementing fixes.

Goal: Capture, classify, and later execute the second real dogfooding feedback batch for the internal TypeDuck-Web playground without reopening completed M24 rows or weakening oracle-backed compatibility claims.

Architecture: M25 treats apps/yune-web/ as the browser dogfooding surface and keeps the browser app on Vite + React + Tailwind CSS + small local components. Browser and UI defects are fixed with Playwright evidence from the local /web/ app; engine-output or ranking defects require pinned TypeDuck v1.1.2 or upstream 1.17.0 fixtures before changing engine behavior. TypeDuck-Web source edits must be regenerated into the tracked patch before they count as landed.

Tech Stack: TypeDuck-Web React/TypeScript, Vite/Bun, Tailwind CSS, small local React components, Playwright, @yune-ime/typeduck-runtime, yune-rime-api C ABI, yune-core, TypeDuck v1.1.2 oracle fixtures, upstream librime 1.17.0 oracle fixtures.

Scope

M25 is the second dogfooding and hardening loop for the internal browser playground:

  • In scope: manual play feedback from http://localhost:5173/web/, first-run and reload behavior, schema switching, settings ergonomics, candidate panel layout, dictionary/detail panel readability, input scenarios, inspector/status usefulness, and local dogfood UI polish.
  • In scope with oracle evidence: any change to candidate text, candidate set, candidate order, segmentation, correction, prediction, reverse lookup, dictionary lookup payloads, or commit behavior.
  • In scope for frontend stack maintenance: keep the dogfood demo on Vite + React + Tailwind CSS + small local components only. Do not add DaisyUI back and do not add another UI framework.
  • Out of scope: editing the separately cloned or deployed TypeDuck-HK/TypeDuck-Web product, treating typeduck.hk/web as the hard oracle, broad design-system work, widening the default RimeApi, or adding unsupported controls that only appear to work.

Relationship To M24

M24 is closed and archived at docs/plans/completed/m24-plan-typeduck-web-dogfooding.md. Do not edit M24 rows for new feedback.

Use M24 as the baseline for:

  • local Tailwind-only component stack,
  • apps/yune-web/e2e/results/m24-dogfooding/ as historical evidence only,
  • TypeDuck-Web patch discipline,
  • the jigaajiusihaa TypeDuck v1.1.2 ordering fixture,
  • the menu/page_size customize key,
  • the Jyutping schema's Mandarin-pinyin affix path `p... as historical M24 behavior; M25-DOGFOOD-08 reopens that dogfood UX and requests bare ` for the luna_pinyin reverse-lookup trigger inside jyut6ping3.

M25 evidence belongs under apps/yune-web/e2e/results/m25-dogfooding/<issue-id>/.

Classification Rules

Classify every report before editing code:

Classification Use when First evidence
Browser integration Worker/runtime/assets/settings wiring fails or drifts from the intended browser contract. Browser console logs, worker diagnostics, state JSON, screenshot.
UI polish The rendered app is confusing, cramped, mislabeled, inaccessible, or inefficient, but engine output is correct. Screenshot plus the exact interaction path.
Engine correctness Candidate output, ranking, commit text, segmentation, correction, prediction, or reverse lookup seems wrong. Pinned oracle fixture or a row marked blocked until fixture capture.
Unsupported / N/A The report asks for behavior intentionally not exposed in the dogfood playground. Short rationale and, if useful, a UI copy/docs change.
Future product integration The report belongs to the real TypeDuck-Web product, not the internal Yune harness. Product-track note; do not edit apps/yune-web/source/ unless the harness also needs it.
Needs triage The symptom is not reproducible or does not yet identify the layer. Screenshot/state capture and a narrow reproduction attempt.

Evidence Rules

  • Save browser evidence under apps/yune-web/e2e/results/m25-dogfooding/<issue-id>/.
  • For every browser-visible fix, capture a screenshot or JSON/state snapshot from the real local /web/ app.
  • For every engine-output fix, add or extend a pinned oracle fixture before implementation.
  • Do not use https://www.typeduck.hk/web/ as a hard oracle. It is a useful feel target only.
  • Keep completed M9/M13/M16/M20/M22/M24 gates green unless a row explicitly changes a supported contract with fresh evidence.

Review Summary For Claude

Use this section as the high-level review map. The running ledger below remains the source of truth for exact evidence paths, owning files, and close criteria.

Execution group Rows Why this grouping matters
Performance and responsiveness blockers M25-DOGFOOD-01, M25-DOGFOOD-03 These decide whether the app feels usable at all. Review first for measurement quality, startup phase attribution, and whether normal typing is separated from deploy/customize/loading work.
Candidate/settings correctness M25-DOGFOOD-02, M25-DOGFOOD-08 These affect user-visible IME behavior. M25-DOGFOOD-02 must fix both the settings key path and rendered candidate cap; M25-DOGFOOD-08 intentionally gives bare ` to the luna_pinyin affix translator inside jyut6ping3, deletes the vestigial bare-grave reverse_lookup slot unless a real consumer is proven, and moves useful secondary lookups into a v-prefixed namespace such as `vl / `vc.
Top-control and settings layout M25-DOGFOOD-04, M25-DOGFOOD-05, M25-DOGFOOD-06, M25-DOGFOOD-10 These should be implemented together only after the correctness blockers are understood, because they share App.tsx, Toolbar.tsx, SchemaSwitcher.tsx, and Preferences.tsx layout surfaces.
Control affordance polish M25-DOGFOOD-07, M25-DOGFOOD-09 These are local component presentation changes. They can be batched if tests preserve the individual checkbox and radio close criteria.

Claude review should focus on four questions:

  1. Is the startup plan still aimed at the validated P0 suspects: debug browser WASM and forced schema invalidation before lower-level micro-optimization?
  2. Does the page-size row protect both menu/page_size browser/runtime wiring and the candidate panel's rendered row cap?
  3. Is the reverse-lookup row clear that bare ` is reserved for the luna_pinyin affix translator inside jyut6ping3, that the current bare-grave reverse_lookup tag is vestigial unless proven otherwise, and that retained secondary lookups need explicit non-bare triggers plus visible UI copy?
  4. Can the UI rows be batched by touched surfaces without losing per-row browser evidence and close criteria?

Patch-Layer Rule

apps/yune-web/source/ is gitignored in the Yune repository. Local edits there are allowed for development, but a M25 row is not closed until the matching tracked artifacts are updated.

Before closing any row that changes TypeDuck-Web source:

  1. Regenerate apps/yune-web/patches/yune-web-runtime.patch from the patched upstream checkout.

  2. Reverse-check from apps/yune-web/source/:

    git apply --reverse --check ..\patches\yune-web-runtime.patch

  3. Forward-check the patch on a clean source checkout reset to apps/yune-web/yune-web.lock.json.

  4. Stage only the tracked artifacts for the slice: the patch, Yune-owned integration files, Playwright tests/evidence, Rust/runtime files, docs, and lock metadata when the upstream source pin changes.

Running Issue Ledger

M25 intake began on 2026-06-21 with user-reported browser dogfooding regressions. Keep adding one row per distinct user-visible symptom using the next M25-DOGFOOD-XX id.

ID Status Classification User-visible issue First repro / evidence Owning surfaces to inspect first Close criteria
M25-DOGFOOD-01 Open Browser integration / performance Refreshing http://localhost:5173/web/ still leaves 載入中 Loading... visible for too long; the user sees no practical improvement compared with the pre-M24 app and considers this unacceptable for real users. Reproduced in the in-app browser on 2026-06-21: reload took 47.752s; startup marker reported totalMs=47331, with runtime:initialized consuming nearly the entire delay after assets loaded. Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-01/reload-startup-repro.json. apps/yune-web/source/src/worker.ts, apps/yune-web/source/src/yune-integration/adapter.ts, apps/yune-web/source/src/yune-integration/assets.ts, packages/yune-typeduck-runtime/src/typeduck.ts, packages/yune-typeduck-runtime/src/module.ts, packages/yune-typeduck-runtime/src/filesystem.ts, crates/yune-rime-api/src/typeduck_web.rs, schema deploy/init paths under crates/yune-rime-api/src/. Close only after a measured startup optimization lands. Add a Playwright startup budget test that records phase timings, preserve startup:complete diagnostics, and capture before/after evidence under this issue id. Target: interactive shell visible quickly and warm reload IME readiness materially below the current ~47s baseline; any chosen budget must be written into the test and justified by local evidence.
M25-DOGFOOD-02 Open Browser integration / settings and candidate pagination The page-size control is hard to find after M24 UI changes, its allowed range is wrong for the requested behavior, and changing it still does not cap the rendered candidate list. The user expects a visible slider/control allowing 3-10 candidates, where setting 9 shows exactly 9 visible candidates. User screenshot showed hai rendering far more than 10 rows. Reproduced in the in-app browser on 2026-06-21: DOM had page-size slider min=4, max=10, value=6, but typing hai rendered 50 visible candidate rows. Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-02/page-size-hai-repro.json. apps/yune-web/source/src/Preferences.tsx, apps/yune-web/source/src/Inputs.tsx, apps/yune-web/source/src/CandidatePanel.tsx, apps/yune-web/source/src/App.tsx, apps/yune-web/source/src/yune-integration/adapter.ts, crates/yune-rime-api/src/typeduck_web.rs, crates/yune-rime-api/src/context_api.rs, crates/yune-rime-api/tests/typeduck_web.rs, apps/yune-web/e2e/yune-typeduck.spec.ts. Close when the UI exposes an obvious 3-10 page-size control and the rendered candidate panel never exceeds the configured page size. Add/extend native typeduck_web coverage for menu/page_size at 3 and 9, add Playwright coverage that sets 3/9/10 and types hai, verify page navigation still works, regenerate and reverse/forward check apps/yune-web/patches/yune-web-runtime.patch, and capture browser JSON/screenshot evidence under this issue id.
M25-DOGFOOD-03 Open Browser integration / typing responsiveness Typing in the textbox can still show the global 載入中 Loading... state and can stall for about a second when entering a character. This makes the dogfood IME feel blocked even after the page becomes visible. User-reported during manual dogfooding on 2026-06-21. Evidence placeholder with exact report: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-03/typing-latency-user-report.json. Needs measured browser repro with keydown-to-candidate timing. apps/yune-web/source/src/CandidatePanel.tsx, apps/yune-web/source/src/rime.ts, apps/yune-web/source/src/worker.ts, apps/yune-web/source/src/App.tsx, apps/yune-web/source/src/Toolbar.tsx, apps/yune-web/source/src/yune-integration/adapter.ts, packages/yune-typeduck-runtime/src/typeduck.ts, crates/yune-rime-api/src/typeduck_web.rs, crates/yune-core/src/engine.rs. Close only after per-key latency is measured and improved. Add Playwright instrumentation for keydown-to-candidate update latency on hai, nei, and a long phrase; split queue wait, worker roundtrip, Rust process_key, and React render time; remove global loading from normal per-key composition; ensure typing remains responsive while startup/deploy/customize is in flight; save before/after latency JSON under this issue id.
M25-DOGFOOD-04 Open UI polish / schema switcher layout The schema list sits in its own vertical block below the three mode buttons, wasting vertical space. The user wants the schema list moved into the same top row as the three buttons. The luna_pinyin option is also currently shown as 普通話, but the user expects the schema name 朙月拼音. User-reported during manual dogfooding on 2026-06-21. Local verification found upstream 1.17.0 luna_pinyin.schema.yaml uses 朙月拼音, while the current TypeDuck-Web source and TypeDuck v1.1.2 captured schema use 普通話. Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-04/schema-switcher-toolbar-and-luna-name-report.json. apps/yune-web/source/src/App.tsx, apps/yune-web/source/src/Toolbar.tsx, apps/yune-web/source/src/SchemaSwitcher.tsx, apps/yune-web/source/src/consts.ts, apps/yune-web/source/src/Inputs.tsx, apps/yune-web/e2e/yune-typeduck.spec.ts. Close when the schema selector is part of the top control row with the three mode buttons on desktop, wraps within the same compact control band on narrow screens, and luna_pinyin is labeled 朙月拼音 as the primary user-facing schema name unless the row is explicitly revised with TypeDuck-v1.1.2-specific rationale. Add Playwright evidence covering desktop and mobile layout, active schema switching, and the visible Luna label; regenerate and reverse/forward check apps/yune-web/patches/yune-web-runtime.patch if source changed.
M25-DOGFOOD-05 Open UI polish / top controls layout The 倉頡版本 Cangjie version control currently lives in a separate lower Web Frontend Controls section, but it is tightly related to schema choice. The user wants it moved to the top beside the schema selection so the page does not spend vertical space on a one-control section. User-reported during manual dogfooding on 2026-06-21. Local verification found the control in apps/yune-web/source/src/Preferences.tsx under Web Frontend Controls, using prefs.isCangjie5 with 三代 Version 3 and 五代 Version 5 segments. Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-05/cangjie-version-top-controls-report.json. apps/yune-web/source/src/App.tsx, apps/yune-web/source/src/Toolbar.tsx, apps/yune-web/source/src/SchemaSwitcher.tsx, apps/yune-web/source/src/Preferences.tsx, apps/yune-web/source/src/Inputs.tsx, apps/yune-web/e2e/yune-typeduck.spec.ts. Close when the Cangjie version segmented control is colocated with the schema selector in the top control band on desktop and wraps within that same compact band on narrow screens, the lower Web Frontend Controls section is removed if it has no remaining controls, and changing 三代/五代 still updates the isCangjie5 customize path. Add Playwright evidence for layout and both Cangjie version values; regenerate and reverse/forward check apps/yune-web/patches/yune-web-runtime.patch if source changed.
M25-DOGFOOD-06 Open UI polish / settings section order The 顯示設定 Display controls section is currently in the second-row right column. The user wants it in the top-row right column, meaning 顯示設定 Display controls and 即時狀態 Live session controls should exchange positions. User-reported during manual dogfooding on 2026-06-21. Local verification found Preferences.tsx renders 即時狀態 Live session controls before 顯示設定 Display controls, causing Display controls to land below the top-right slot in the two-column settings grid. Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-06/display-live-section-order-report.json. apps/yune-web/source/src/Preferences.tsx, apps/yune-web/source/src/Inputs.tsx, apps/yune-web/e2e/yune-typeduck.spec.ts. Close when Display controls render before Live session controls in the settings grid, occupy the top-right column on desktop, and Live session controls move to the position Display controls previously occupied. Preserve all existing control labels, state bindings, and live-option behavior. Add Playwright desktop and narrow-viewport evidence for the section order and bounding boxes; regenerate and reverse/forward check apps/yune-web/patches/yune-web-runtime.patch if source changed.
M25-DOGFOOD-07 Open UI polish / binary control affordance Binary settings are currently shown as rounded rectangle pill switches that fill blue when checked. The user finds this unintuitive and wants these binary toggles to be raw checkbox-style controls instead. User-reported during manual dogfooding on 2026-06-21 with screenshot codex-clipboard-da4df560-7b7a-49fd-814f-4cfe5cbd6968.png. Local verification found Toggle renders input type="checkbox" with className="yd-switch" in Inputs.tsx, and .yd-switch applies the pill styling in index.css; .yd-check already exists as a square checkbox style. Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-07/raw-checkbox-binary-controls-report.json. apps/yune-web/source/src/Inputs.tsx, apps/yune-web/source/src/index.css, apps/yune-web/source/src/App.tsx, apps/yune-web/source/src/ThemeSwitcher.tsx, apps/yune-web/source/src/Preferences.tsx, apps/yune-web/e2e/yune-typeduck.spec.ts. Close when binary dogfood controls render as checkbox-style controls rather than pill switches, checked/unchecked states remain clear and keyboard-accessible, and no settings-panel binary control uses the blue rounded-pill affordance. Decide explicitly whether the app theme toggle remains a specialized icon switch or also moves to checkbox styling. Add Playwright evidence for checked and unchecked states across engine/session/display/frontend settings plus the inspector toggle; regenerate and reverse/forward check apps/yune-web/patches/yune-web-runtime.patch if source changed.
M25-DOGFOOD-08 Open Browser integration / reverse lookup schema trigger Mandarin luna_pinyin reverse lookup inside the jyut6ping3 schema currently appears to require `p. The user clarified that the single grave accent ` should trigger the luna_pinyin affix translator. The current bare-grave reverse_lookup tag is vestigial in the desktop schema because it has no translator/dictionary, so it should be deleted unless a real consumer is proven. Useful secondary lookup paths such as Loengfan and Cangjie should move out of the bare/Mandarin namespace, preferably to v-prefixed triggers such as `vl and `vc; avoid single-letter prefixes that shadow normal pinyin initials. The web UI must explicitly show which trigger maps to which lookup/schema. While trying reverse lookup, the app is also slow and can show 執行操作時發生錯誤。如輸入法不能正常運作，請重新載入頁面。 / An error occurred while performing the operation...; treat that performance/error symptom as related to M25-DOGFOOD-01/M25-DOGFOOD-03 unless a fresh repro proves a distinct reverse-lookup crash. User-reported during manual dogfooding on 2026-06-21, then clarified after external review on 2026-06-21. Local verification found apps/yune-web/source/schema/jyut6ping3.schema.yaml has a current vestigial bare-grave reverse_lookup tag with no translator/dictionary, while functional luna_pinyin, loengfan, and cangjie lookups use `p, `l, and `c respectively; jyut6ping3_mobile.schema.yaml only exposes the current `p luna_pinyin path. The TypeDuck v1.1.2 captured schema also has `p, so this is a documented M25 web-profile divergence rather than an oracle match. Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-08/reverse-lookup-bare-grave-trigger-report.json. apps/yune-web/source/schema/jyut6ping3.schema.yaml, apps/yune-web/source/schema/jyut6ping3_mobile.schema.yaml, apps/yune-web/source/src/consts.ts, apps/yune-web/source/src/SchemaSwitcher.tsx, crates/yune-rime-api/src/schema_install.rs, crates/yune-rime-api/tests/typeduck_web.rs, apps/yune-web/e2e/yune-typeduck.spec.ts. Close when the web Jyutping schema uses bare ` for luna_pinyin reverse lookup; the vestigial bare-grave reverse_lookup tag is removed or justified with a concrete non-bare prefix; retained Loengfan/Cangjie side lookups use explicit non-bare triggers such as `vl / `vc; browser evidence shows typing `zhe reaches luna_pinyin without needing p; native/browser tests prove pinyin initials such as lai, ci, xi, and re are not stolen by secondary lookups; and the web UI lists the lookup triggers per active schema. Document the deliberate divergence from TypeDuck v1.1.2 `p behavior, update visible hints, capture browser JSON/screenshot evidence, and regenerate plus reverse/forward check apps/yune-web/patches/yune-web-runtime.patch if source changed.
M25-DOGFOOD-09 Open UI polish / control affordance The 候選排版 Candidate Menu Layout setting currently uses a segmented control. The user wants it changed to a radio selection. User-reported during manual dogfooding on 2026-06-21. Local verification found Preferences.tsx maps CANDIDATE_MENU_LAYOUT_LABELS into Segment controls for prefs.candidateMenuLayout; Inputs.tsx already provides a Radio component. Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-09/candidate-layout-radio-report.json. apps/yune-web/source/src/Preferences.tsx, apps/yune-web/source/src/Inputs.tsx, apps/yune-web/source/src/consts.ts, apps/yune-web/e2e/yune-typeduck.spec.ts. Close when Candidate Menu Layout renders as radio choices, not a segmented group; both horizontal and vertical options remain labeled, keyboard-accessible, and bound to candidateMenuLayout; switching options still changes the candidate panel layout. Add Playwright evidence for selecting both options and capturing the resulting panel layout; regenerate and reverse/forward check apps/yune-web/patches/yune-web-runtime.patch if source changed.
M25-DOGFOOD-10 Open UI polish / layout alignment The 輸入法設定 IME Settings section appears to use a different horizontal margin than the playground section above it. The user wants the settings block aligned with the content above so the page columns share the same left/right edges. User-reported during manual dogfooding on 2026-06-21. Local verification found App.tsx renders the main playground inside <main className="m-auto p-8 max-w-7xl">, while Preferences.tsx wraps IME Settings in a nested <section className="mx-auto max-w-6xl">, making settings narrower than the preceding content stack. Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-10/ime-settings-margin-alignment-report.json. apps/yune-web/source/src/App.tsx, apps/yune-web/source/src/Preferences.tsx, apps/yune-web/source/src/index.css, apps/yune-web/e2e/yune-typeduck.spec.ts. Close when the IME Settings heading and settings grid align horizontally with the content above on desktop and narrow viewports. Add Playwright evidence comparing bounding boxes for the playground content and IME Settings wrapper; left and right edge differences should be within 2 px unless a documented scrollbar/device-pixel reason requires a slightly wider tolerance. Preserve the settings grid behavior from M25-DOGFOOD-06 and regenerate plus reverse/forward check apps/yune-web/patches/yune-web-runtime.patch if source changed.

Accepted Review Corrections

  • M25-DOGFOOD-01 startup work must address the two validated P0 suspects before broader micro-optimization: the browser WASM artifact is currently built from target/wasm32-unknown-emscripten/debug, and the TypeDuck-Web adapter forces deploy freshness by calling invalidateDeployedSchema(...) before currentRuntime.deploy(). The review evidence is tracked at apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-01/claude-review-p0-notes.json.
  • Do not treat fast asset fetches as proof that startup is network-bound. The current reproduction showed assets loading in under 100 ms while runtime:initialized consumed about 47 seconds, so Task 2 must re-measure release WASM and deploy reuse before setting the final warm-reload budget.
  • M25-DOGFOOD-08 follow-up review found that desktop jyut6ping3's current bare-grave reverse_lookup tag has no translator/dictionary and is not a functional Jyutping lookup. The M25 decision is therefore: bare ` belongs to luna_pinyin; useful secondary lookups such as Loengfan/Cangjie should move to an explicit non-bare namespace, preferably v-prefixed triggers such as `vl and `vc to avoid shadowing normal pinyin initials; the web UI must show which trigger maps to which lookup.

Closeout Evidence

All M25 rows closed on 2026-06-21. The intake rows above are preserved as the original report text; this closeout table is the authoritative status, owning-test, and evidence index for the completed ledger.

Row Status Owning tests Evidence
M25-DOGFOOD-01 Closed M25 DOGFOOD-01 startup uses release wasm and records deploy cache reuse; npm.cmd --prefix apps/yune-web/source run build apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-01/startup-after.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-01/screenshot-startup-ready.png
M25-DOGFOOD-02 Closed typeduck_adapter_real_assets_page_size_customize_limits_context_page; M25 DOGFOOD-02 page-size slider caps rows and preserves candidate paging apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-02/page-size-summary.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-02/screenshot-page-size-9-hai.png
M25-DOGFOOD-03 Closed M25 DOGFOOD-03 typing stays out of global loading and records key latency apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-03/typing-latency-after.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-03/screenshot-typing-responsive.png
M25-DOGFOOD-04 Closed M25 DOGFOOD-04 schema switcher shares the top control row apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-04/schema-switcher-toolbar-desktop.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-04/screenshot-schema-switcher-toolbar-desktop.png; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-04/screenshot-schema-switcher-toolbar-mobile.png
M25-DOGFOOD-05 Closed M25 DOGFOOD-05 Cangjie version control lives in the top control band apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-05/cangjie-version-top-controls.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-05/screenshot-cangjie-version-top-controls.png
M25-DOGFOOD-06 Closed M25 DOGFOOD-06 display controls precede live session controls apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-06/display-live-section-order-desktop.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-06/screenshot-display-live-section-order-desktop.png; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-06/screenshot-display-live-section-order-mobile.png
M25-DOGFOOD-07 Closed M25 DOGFOOD-07 binary controls use checkbox affordance apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-07/checkbox-affordance-summary.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-07/screenshot-checkbox-affordance-desktop.png; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-07/screenshot-checkbox-affordance-mobile.png
M25-DOGFOOD-08 Closed typeduck_adapter_browser_app_assets_load_jyutping_mandarin_pinyin_reverse_lookup; M25 DOGFOOD-08 Jyutping bare grave routes to Luna reverse lookup apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-08/reverse-lookup-summary.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-08/bare-grave-zhe-state.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-08/screenshot-bare-grave-zhe.png
M25-DOGFOOD-09 Closed M25 DOGFOOD-09 candidate menu layout uses radio choices apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-09/candidate-layout-horizontal.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-09/candidate-layout-vertical.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-09/screenshot-candidate-layout-horizontal.png; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-09/screenshot-candidate-layout-vertical.png
M25-DOGFOOD-10 Closed M25 DOGFOOD-10 IME settings align with the playground content edges apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-10/settings-alignment-desktop.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-10/settings-alignment-mobile.json; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-10/screenshot-settings-alignment-desktop.png; apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-10/screenshot-settings-alignment-mobile.png

Patch discipline evidence: apps/yune-web/patches/yune-web-runtime.patch was regenerated from apps/yune-web/source/ and checked with both reverse and forward git apply --check paths after source changes.

Intake Task

Task 1: Convert User Feedback Into M25 Rows

Files:

  • Modify: docs/plans/m25-plan-typeduck-web-dogfooding-round-2.md

  • Create evidence as needed: apps/yune-web/e2e/results/m25-dogfooding/<issue-id>/

  • Step 1: Split the feedback into distinct symptoms

    Treat each independently reproducible behavior as one row. If two comments share one root cause but have different user-visible symptoms, keep separate rows until triage proves they should close together.

  • Step 2: Assign stable ids

    Use M25-DOGFOOD-01, M25-DOGFOOD-02, and so on. Do not renumber rows after they are referenced by evidence, commits, or tests.

  • Step 3: Classify each row before implementation

    Use one of the classifications in this plan. Mark ambiguous reports as Needs triage; do not force them into browser or engine buckets prematurely.

  • Step 4: Name the first evidence path

    For browser reports, record the intended evidence directory, for example:

    apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-01/

    For engine reports, record the fixture family first: typeduck-v1.1.2 for TypeDuck profile behavior or upstream-1.17.0 for default upstream behavior.

  • Step 5: Name the owning surfaces

    List the files or test families to inspect first. For TypeDuck-Web source changes, include the source/ file and the tracked patch requirement.

  • Step 6: Write close criteria that can be verified

    Each row needs concrete close criteria: owning test, evidence path, patch regeneration if applicable, and the focused gate to run.

Task 2: M25-DOGFOOD-01 Startup Performance Optimization

Files:

  • Modify: Cargo.toml

  • Modify: scripts/typeduck-wasm-build.sh

  • Modify: apps/yune-web/source/src/worker.ts

  • Modify: apps/yune-web/source/src/yune-integration/adapter.ts

  • Modify if profiling still points there: packages/yune-typeduck-runtime/src/typeduck.ts

  • Modify if profiling still points there: packages/yune-typeduck-runtime/src/filesystem.ts

  • Modify if profiling still points there: crates/yune-rime-api/src/typeduck_web.rs

  • Test: apps/yune-web/e2e/yune-typeduck.spec.ts

  • Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-01/

  • Step 1: Preserve the failing startup characterization and artifact facts

    Keep the current reload evidence as the baseline, then add a Playwright startup budget characterization that reloads /web/, waits for document.documentElement.dataset.yuneLoading === "false", captures the latest startup:complete diagnostic, and writes m25-dogfooding/M25-DOGFOOD-01/startup-before.json. Also record the shipped .wasm size, whether the build came from debug or release, and whether Emscripten/wasm optimization flags were applied. The first test may fail against the current ~47s local baseline, but the final budget should be chosen after the P0 fixes below are measured.

  • Step 2: Ship release-mode browser WASM first

    Update the TypeDuck-Web build path so the browser artifact is not copied from target/wasm32-unknown-emscripten/debug. Add conservative release tuning in Cargo.toml for the existing workspace (opt-level = 3, lto = true, codegen-units = 1, panic = "abort", and strip = true if supported by the active toolchain), update scripts/typeduck-wasm-build.sh to build typeduck_web_module with --release, copy from target/wasm32-unknown-emscripten/release, and add Emscripten optimization flags such as -O3 or -Oz to the link path. Run wasm-opt -O3 only when available and record whether it was skipped. Preserve the existing JS glue contract: createYuneTypeduckModule, cwrap, UTF8ToString, FS, and IDBFS must still exist after the build.

  • Step 3: Stop rebuilding/deploying fresh state on every reload

    Remove or guard the unconditional invalidateDeployedSchema(currentFs, currentPrepareOptions) call before currentRuntime.deploy(). Do not simply delete the safety check; replace it with freshness validation based on assetVersion, schema/resource checksums, or another deterministic deploy stamp. Persist and reuse /rime/build through IDBFS when fresh. If jyut6ping3_mobile still spends startup time generating stable table/reverse artifacts, ship or prefer precompiled artifacts only after adding fixture-backed evidence that the resulting candidates and paging behavior remain unchanged.

  • Step 4: Re-measure after the P0 fixes before choosing deeper work

    Rebuild the patchable TypeDuck-Web source, reload /web/, and save startup-after-release-wasm.json and startup-after-deploy-reuse.json as applicable. Compare startup:complete.totalMs, the user-visible loading duration, .wasm size, and deploy/cache diagnostics against reload-startup-repro.json. The final warm-reload budget must be written from this measured optimized baseline, not guessed from the pre-fix 47s number.

  • Step 5: Add finer runtime:initialized phase markers only if needed

    If release WASM plus deploy reuse do not meet the startup budget, split runtime:initialized into nested markers around TypeDuckRuntime.init, filesystem mount/sync, schema deploy, default schema selection, dictionary/table loading, and persistence sync. The next evidence file must show which sub-phase owns the remaining delay.

  • Step 6: Keep startup latency separate from typing latency

    Startup optimization must not hide the separate typing-stutter problem. If typing is still blocked while startup continues, keep M25-DOGFOOD-03 open and measure it separately instead of claiming startup fixed the perceived performance problem.

  • Step 7: Prove the improvement and regenerate the TypeDuck-Web patch if source changed

    Re-run the startup test, save the final startup-after.json, and compare startup:complete.totalMs plus the visible loading duration against the baseline evidence. The final row update must state before/after timings, the release/deploy evidence paths, and any remaining bottleneck. If any file under apps/yune-web/source/ changed, regenerate apps/yune-web/patches/yune-web-runtime.patch, reverse-check it from source/, and forward-check it on a clean source checkout.

Task 3: M25-DOGFOOD-02 Page-Size Slider And Candidate Cap

Files:

  • Modify: apps/yune-web/source/src/Preferences.tsx

  • Modify if needed: apps/yune-web/source/src/Inputs.tsx

  • Modify if needed: apps/yune-web/source/src/CandidatePanel.tsx

  • Modify if needed: apps/yune-web/source/src/yune-integration/adapter.ts

  • Modify if needed: crates/yune-rime-api/src/typeduck_web.rs

  • Test: crates/yune-rime-api/tests/typeduck_web.rs

  • Test: apps/yune-web/e2e/yune-typeduck.spec.ts

  • Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-02/

  • Step 1: Add failing native page-size coverage for 3 and 9

    Extend typeduck_adapter_real_assets_page_size_customize_limits_context_page or add a focused companion test that customizes menu/page_size to 3 and 9, types a high-candidate input such as hai, and asserts context.page_size and context.candidates.len() match the requested cap.

  • Step 2: Add failing browser coverage for the visible control

    Add a Playwright test that locates the page-size control by its Cantonese/English label, asserts the range is 3 through 10, sets it to 9, types hai, and asserts exactly 9 visible candidate rows. Repeat with 3 to protect the lower bound.

  • Step 3: Fix discoverability and the allowed range

    Keep the control in the settings UI, but make it visibly discoverable in the candidate/display area and set its range to 3..10. The visible value must update immediately when the slider changes.

  • Step 4: Fix the runtime cap

    Trace the browser response after Rime.customize({ pageSize }). If native typeduck_web returns a paged context.candidates array but the browser renders all candidates, fix CandidatePanel or the adapter mapping. If native returns too many candidates for menu/page_size, fix typeduck_web.rs or context_api.rs. Do not hide extra rows with CSS while leaving selection/page navigation semantics wrong.

  • Step 5: Prove paging and selection still work

    The browser test must verify first-page row count, next/previous page behavior, and digit selection after the cap is applied. Save JSON and screenshots under M25-DOGFOOD-02.

  • Step 6: Regenerate the TypeDuck-Web patch if source changed

    If any file under apps/yune-web/source/ changed, regenerate apps/yune-web/patches/yune-web-runtime.patch, reverse-check it from source/, and forward-check it on a clean source checkout.

Task 4: M25-DOGFOOD-03 Typing Responsiveness And Loading-State Separation

Files:

  • Modify: apps/yune-web/source/src/CandidatePanel.tsx

  • Modify: apps/yune-web/source/src/rime.ts

  • Modify: apps/yune-web/source/src/worker.ts

  • Modify: apps/yune-web/source/src/App.tsx

  • Modify: apps/yune-web/source/src/Toolbar.tsx

  • Modify if profiling points there: apps/yune-web/source/src/yune-integration/adapter.ts

  • Modify if profiling points there: packages/yune-typeduck-runtime/src/typeduck.ts

  • Modify if profiling points there: crates/yune-rime-api/src/typeduck_web.rs

  • Test: apps/yune-web/e2e/yune-typeduck.spec.ts

  • Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-03/

  • Step 1: Add keypress latency instrumentation

    Add temporary or permanent diagnostics that record keydown time, Rime.processKey queued time, worker receive time, Rust response time, page receive time, and candidate-panel render time. Save a browser JSON baseline for h, ha, hai, nei, and jigaajiusihaa under M25-DOGFOOD-03.

  • Step 2: Add a failing Playwright responsiveness test

    Add a Playwright test that types into the real textarea using actual keypresses and asserts an explicit latency budget for keydown-to-candidate update. The first version should fail or record the current bad p95/p99 latency so the implementation has a hard target.

  • Step 3: Separate global loading state from normal composition

    載入中 Loading... should mean startup, schema deploy, or settings redeploy, not ordinary per-key composition. Split loading state in App.tsx/Toolbar.tsx/rime.ts so normal processKey, stageAi, page flip, and candidate selection do not keep the global loading indicator active. If an action needs a local pending state, show it near that control instead of blocking the textbox.

  • Step 4: Avoid queueing keystrokes behind long non-key actions

    Inspect the rime.ts single-message queue. If processKey messages wait behind startup/customize/deploy, split high-priority key events from low-priority settings work, coalesce stale option/customize calls, or block typing until the IME is truly ready with an explicit disabled state. Do not let the user type into an apparently live IME while key events are silently delayed.

  • Step 5: Reduce the slowest measured per-key path

    If latency is dominated by Rust process_key, profile yune_typeduck_process_key and candidate serialization. If latency is dominated by rendering, cap rendered rows through M25-DOGFOOD-02 and avoid rebuilding expensive dictionary/detail data for every row. If latency is dominated by AI staging, ensure AI remains second-pass, cancellable/stale-result guarded, and default-off.

  • Step 6: Prove typing responsiveness after the fix

    Re-run the browser responsiveness test and save typing-latency-after.json with before/after p50/p95/p99 timings. The final evidence must show the textbox accepts keypresses without a visible one-second stall and the global loading indicator does not appear for normal composition.

  • Step 7: Regenerate the TypeDuck-Web patch if source changed

    If any file under apps/yune-web/source/ changed, regenerate apps/yune-web/patches/yune-web-runtime.patch, reverse-check it from source/, and forward-check it on a clean source checkout.

Task 5: M25-DOGFOOD-04 Compact Schema Switcher And Luna Label

Files:

  • Modify: apps/yune-web/source/src/App.tsx

  • Modify: apps/yune-web/source/src/Toolbar.tsx

  • Modify: apps/yune-web/source/src/SchemaSwitcher.tsx

  • Modify: apps/yune-web/source/src/consts.ts

  • Modify if segment markup needs shared layout changes: apps/yune-web/source/src/Inputs.tsx

  • Test: apps/yune-web/e2e/yune-typeduck.spec.ts

  • Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-04/

  • Step 1: Preserve the schema-name verification

    Keep schema-switcher-toolbar-and-luna-name-report.json as the issue baseline. The implementation should treat the current web label 普通話 as a UI/source mismatch to resolve, not as proof that the user-facing label is correct. Upstream 1.17.0 names luna_pinyin as 朙月拼音; TypeDuck v1.1.2 names it 普通話, so any decision to keep 普通話 must be documented in this row with explicit TypeDuck-specific rationale.

  • Step 2: Add failing browser layout coverage

    Add a Playwright test that loads /web/ at a desktop viewport and asserts the schema switcher belongs to the same top control band as the ASCII/simplification/full-shape buttons. The test should check that [data-yune-schema-switcher] is inside or directly owned by the toolbar/top-controls component, the top edges of the three mode buttons and schema segments are aligned within one compact row, and no separate standalone schema block appears between the toolbar and status strip. Add a narrow viewport assertion that the same controls wrap within one compact band instead of becoming a separate full-width section.

  • Step 3: Move the schema switcher into the toolbar row

    Refactor Toolbar.tsx to accept activeSchema and setActiveSchema, render SchemaSwitcher next to the three mode buttons, and keep the loading indicator in the same top control area without pushing the textarea downward. Remove the standalone <SchemaSwitcher ... /> call from App.tsx. Keep accessible labels for each schema option and preserve the existing data-yune-schema-switcher hook for tests.

  • Step 4: Make the Luna label explicit

    Update SCHEMA_OPTIONS in consts.ts so luna_pinyin uses 朙月拼音 as the primary visible schemaName. Keep Luna Pinyin or Mandarin only as a secondary label or tooltip if useful. Do not change engine output/ranking behavior for this UI-only row.

  • Step 5: Prove schema switching still works

    In the browser test, switch from jyut6ping3_mobile to luna_pinyin and back, assert the visible active option changes, assert the Luna option text includes 朙月拼音, and type a short smoke input after switching to prove the selector still drives the worker path.

  • Step 6: Capture visual evidence

    Save desktop and narrow-viewport screenshots plus a JSON summary under M25-DOGFOOD-04. The JSON should include viewport size, active schema before/after switching, bounding boxes for the three mode buttons and schema switcher, and the visible text for the Luna option.

  • Step 7: Regenerate the TypeDuck-Web patch if source changed

    If any file under apps/yune-web/source/ changed, regenerate apps/yune-web/patches/yune-web-runtime.patch, reverse-check it from source/, and forward-check it on a clean source checkout.

Task 6: M25-DOGFOOD-05 Cangjie Version In Top Controls

Files:

  • Modify: apps/yune-web/source/src/App.tsx

  • Modify: apps/yune-web/source/src/Toolbar.tsx

  • Modify: apps/yune-web/source/src/SchemaSwitcher.tsx if the schema row becomes the shared top-control layout

  • Modify: apps/yune-web/source/src/Preferences.tsx

  • Modify if segment markup needs shared compact styling: apps/yune-web/source/src/Inputs.tsx

  • Test: apps/yune-web/e2e/yune-typeduck.spec.ts

  • Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-05/

  • Step 1: Preserve the Cangjie-version baseline

    Keep cangjie-version-top-controls-report.json as the issue baseline. The current control lives in Preferences.tsx under Web Frontend Controls, uses prefs.isCangjie5, and exposes 三代 Version 3 plus 五代 Version 5. The implementation must move this existing control, not create a second independent setting.

  • Step 2: Add failing browser layout coverage

    Add a Playwright test that loads /web/ at a desktop viewport and asserts the Cangjie version control is in the same top control band as [data-yune-schema-switcher]. The test should assert that the 三代 Version 3 and 五代 Version 5 controls appear before the status strip and textarea, and that no lower section titled Web Frontend Controls remains when this was its only control. Add a narrow viewport assertion that schema selection and Cangjie version wrap within the same compact top band.

  • Step 3: Move the control into the top control component

    Pass isCangjie5 and setIsCangjie5 from App.tsx into the top toolbar/control component created or updated by M25-DOGFOOD-04. Render a compact segmented control next to the schema selector with the label 倉頡版本 Cangjie version, options 三代 Version 3 and 五代 Version 5, and the existing Segment interaction pattern. Keep the control keyboard-accessible and do not add another UI framework.

  • Step 4: Remove the empty lower section

    Delete the Web Frontend Controls section from Preferences.tsx if moving Cangjie version leaves it empty. If later rows add other frontend controls, keep those controls in an appropriately named settings group, but do not keep a one-control Cangjie section at the bottom.

  • Step 5: Prove the existing customize path still works

    In Playwright, switch to the Cangjie schema, choose 三代 Version 3, apply or trigger the existing settings flow, then choose 五代 Version 5. Capture the active segment state and a worker/debug signal showing isCangjie5 changed, or use the existing UI/evidence hook that proves Rime.customize({ isCangjie5 }) was invoked. Do not change engine behavior or Cangjie candidate ordering in this UI row.

  • Step 6: Capture visual evidence

    Save desktop and narrow-viewport screenshots plus a JSON summary under M25-DOGFOOD-05. The JSON should include viewport size, bounding boxes for schema selector and Cangjie version, active Cangjie version before/after toggling, and whether any lower Web Frontend Controls section remains.

  • Step 7: Regenerate the TypeDuck-Web patch if source changed

    If any file under apps/yune-web/source/ changed, regenerate apps/yune-web/patches/yune-web-runtime.patch, reverse-check it from source/, and forward-check it on a clean source checkout.

Task 7: M25-DOGFOOD-06 Display Controls Above Live Session Controls

Files:

  • Modify: apps/yune-web/source/src/Preferences.tsx

  • Modify if shared section markup needs a test hook: apps/yune-web/source/src/Inputs.tsx

  • Test: apps/yune-web/e2e/yune-typeduck.spec.ts

  • Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-06/

  • Step 1: Preserve the section-order baseline

    Keep display-live-section-order-report.json as the issue baseline. The current Preferences.tsx order is engine/session dogfood controls, 即時狀態 Live session controls, 顯示設定 Display controls, then 網頁前端 Web Frontend Controls. The requested order swaps Display controls and Live session controls without changing the controls inside either section.

  • Step 2: Add failing browser layout coverage

    Add a Playwright test that loads /web/ at a desktop viewport and locates the headings 顯示設定 Display controls and 即時狀態 Live session controls. Assert that Display controls is visually above Live session controls and occupies the top-right settings-grid slot. Use bounding boxes rather than only DOM order, because the user complaint is about the rendered two-column layout. Add a narrow viewport assertion that Display controls still appears before Live session controls in a readable single-column flow.

  • Step 3: Swap the sections in Preferences.tsx

    Move the full 顯示設定 Display controls section above the 即時狀態 Live session controls section. Preserve every existing Display control: display languages, page size, candidate menu layout, Chinese typeface, candidate Jyutping, and reverse code display. Preserve every existing Live session control: ASCII mode, full shape, simplification, traditionalization, extended charset, and disabled.

  • Step 4: Prove state bindings still work

    Extend the browser test to toggle one Display control and one Live session control after the swap. Use existing UI state hooks where available, or assert the active/checked DOM state directly. This prevents a layout-only edit from accidentally detaching prefs.displayLanguages, prefs.pageSize, prefs.isAsciiMode, or related state setters.

  • Step 5: Capture visual evidence

    Save desktop and narrow-viewport screenshots plus a JSON summary under M25-DOGFOOD-06. The JSON should include viewport size, section heading order, bounding boxes for Display controls and Live session controls, and the toggled control state checked in Step 4.

  • Step 6: Regenerate the TypeDuck-Web patch if source changed

    If any file under apps/yune-web/source/ changed, regenerate apps/yune-web/patches/yune-web-runtime.patch, reverse-check it from source/, and forward-check it on a clean source checkout.

Task 8: M25-DOGFOOD-07 Checkbox Affordance For Binary Controls

Files:

  • Modify: apps/yune-web/source/src/Inputs.tsx

  • Modify: apps/yune-web/source/src/index.css

  • Modify: apps/yune-web/source/src/App.tsx

  • Modify if the theme toggle is included in the change: apps/yune-web/source/src/ThemeSwitcher.tsx

  • Test: apps/yune-web/e2e/yune-typeduck.spec.ts

  • Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-07/

  • Step 1: Preserve the binary-control baseline

    Keep raw-checkbox-binary-controls-report.json as the issue baseline. The current shared Toggle component in Inputs.tsx renders input type="checkbox" with className="yd-switch", and .yd-switch in index.css creates the rounded pill switch shown in the user screenshot. .yd-check already exists as the square checkbox styling used by checklist controls.

  • Step 2: Add failing browser affordance coverage

    Add a Playwright test that loads /web/, finds representative binary settings in the preferences area, and asserts they do not use the rounded pill visual dimensions. Cover at least one engine/session toggle, one live-session toggle, one display/settings toggle if present after M25-DOGFOOD-06, and the Yune inspector toggle. The test should record each control's label, checked state, class list, bounding box, and computed border radius.

  • Step 3: Change the shared Toggle presentation to checkbox style

    Update Toggle in Inputs.tsx to render the binary control with the checkbox-style class used by raw checkboxes, or rename the shared class so both Toggle and Checkbox use the same square checkbox affordance. Keep semantic input type="checkbox", label click behavior, checked binding, NO_AUTO_FILL, and focus-ring accessibility. Do not change Radio, Segment, or Range controls.

  • Step 4: Retire or narrow yd-switch usage

    Remove the rounded-pill styling from settings-panel binary controls. Update the Yune inspector toggle in App.tsx to use the same checkbox affordance. Decide explicitly in the row evidence whether ThemeSwitcher.tsx remains a specialized icon switch because it overlays sun/moon icons, or whether it also becomes a checkbox-style control. If the theme toggle remains specialized, rename its class or add a comment so yd-switch no longer means generic binary setting.

  • Step 5: Prove checked and unchecked states remain obvious

    In Playwright, toggle representative controls on and off and assert both states are visible through the checkbox mark/fill, not only through surrounding text. Verify keyboard focus still lands on the input and Space toggles the checked state.

  • Step 6: Capture visual evidence

    Save desktop and narrow-viewport screenshots plus a JSON summary under M25-DOGFOOD-07. The JSON should include the class names and computed styles for each checked control, unchecked control, and any explicitly exempted theme toggle.

  • Step 7: Regenerate the TypeDuck-Web patch if source changed

    If any file under apps/yune-web/source/ changed, regenerate apps/yune-web/patches/yune-web-runtime.patch, reverse-check it from source/, and forward-check it on a clean source checkout.

Task 9: M25-DOGFOOD-08 Bare-Grave Luna Reverse Lookup Trigger And Namespace Reassignment

Files:

  • Modify: apps/yune-web/source/schema/jyut6ping3.schema.yaml

  • Modify: apps/yune-web/source/schema/jyut6ping3_mobile.schema.yaml

  • Modify: apps/yune-web/source/src/consts.ts

  • Modify if trigger copy needs a richer display: apps/yune-web/source/src/SchemaSwitcher.tsx

  • Modify if schema installation rejects the reassigned prefixes: crates/yune-rime-api/src/schema_install.rs

  • Test: crates/yune-rime-api/tests/typeduck_web.rs

  • Test: apps/yune-web/e2e/yune-typeduck.spec.ts

  • Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-08/

  • Step 1: Preserve the trigger baseline and superseded assumption

    Keep reverse-lookup-bare-grave-trigger-report.json as the issue baseline. Current web schemas use luna_pinyin.prefix: "`p" and luna_pinyin: "^`p[a-z']*;?$" under speller/algebra; desktop jyut6ping3.schema.yaml also has reverse_lookup.prefix: ""andreverse_lookup: "^[a-z']*;?$", but that tag has no translator/dictionary and currently produces no useful lookup. Functional secondary lookups are loengfan on `l and cangjie on `c. TypeDuck v1.1.2 captured schema data also uses `p for luna_pinyin. Treat M25 as an intentional web-profile divergence that gives bare ` to luna_pinyin.

  • Step 2: Add failing native/schema coverage

    Add or extend a typeduck_web test that installs the Jyutping web/mobile schema and proves an input beginning with bare grave, for example `zhe, routes into the luna_pinyin affix translator. The test should also prove the vestigial desktop reverse_lookup tag was removed or moved away from bare grave, and that retained secondary lookups use explicit v-prefixed triggers such as `vl for Loengfan and `vc for Cangjie. Add overlap probes such as `lai, `ci, `xi, and `re to prove normal pinyin initials still route to luna_pinyin; if any secondary lookup intentionally uses another prefix, document the dead-zone it creates and justify it in evidence. Record whether `pzhe remains accepted as a compatibility alias or is no longer required.

  • Step 3: Add failing browser coverage

    Add a Playwright test that loads /web/, keeps the Jyutping schema active, types `zhe, and asserts the candidate/debug state identifies the luna_pinyin reverse-lookup path without requiring the p prefix. If Loengfan/Cangjie secondary lookups remain exposed, add browser assertions for their explicit triggers, for example `vl and `vc. Also assert the visible UI explains the trigger map for the active schema. Capture whether the generic operation-error toast appears; if the toast appears because the worker is blocked or slow rather than because of trigger parsing, link that finding back to M25-DOGFOOD-01 and M25-DOGFOOD-03 instead of broadening this row.

  • Step 4: Change the schema trigger

    Update the luna_pinyin affix translator settings in jyut6ping3.schema.yaml and jyut6ping3_mobile.schema.yaml so the prefix is bare ` and the alphabet regex accepts bare-grave Mandarin pinyin input, for example ^`[a-z']*;?$. In desktop jyut6ping3.schema.yaml, remove the vestigial bare-grave reverse_lookup tag unless implementation finds a real translator/dictionary consumer. Move retained Loengfan and Cangjie side lookups out of single-letter pinyin-initial prefixes, preferably to a v namespace such as `vl and `vc. If Yune/librime first-match ordering requires it, place specific v-prefixed affix segmentors before the bare-grave luna_pinyin catch-all and prove the order with the tests from Step 2. Keep any different r/x/v decision explicit in evidence.

  • Step 5: Update user-facing hints

    Update SCHEMA_OPTIONS in consts.ts and, if necessary, SchemaSwitcher.tsx so the web UI is explicit about triggers for each active schema. For jyut6ping3, show bare-grave Mandarin lookup such as `zhe -> 這（普通話 / luna_pinyin 反查） and list retained secondary triggers separately, for example `vl... for Loengfan and `vc... for Cangjie. Do not describe bare ` as a generic Jyutping lookup trigger. If luna_pinyin or Cangjie schemas have their own lookup hints, keep them separate so users can see which trigger belongs to which schema.

  • Step 6: Prove behavior and error handling

    Run the native/schema test and browser test. Save JSON evidence with the typed input, active schema id, reverse-lookup tag/source if available, first candidates, selected secondary namespace, affix segmentor order, visible UI trigger copy, whether the operation-error toast appeared, and per-key timing if available. The evidence must state that this is a deliberate web-profile divergence from TypeDuck v1.1.2's `p luna_pinyin trigger. If reverse lookup remains slow after the trigger fix, leave or cross-link the latency to M25-DOGFOOD-01/M25-DOGFOOD-03; open a new row only if the trigger fix exposes a distinct crash.

  • Step 7: Regenerate the TypeDuck-Web patch if source changed

    If any file under apps/yune-web/source/ changed, regenerate apps/yune-web/patches/yune-web-runtime.patch, reverse-check it from source/, and forward-check it on a clean source checkout.

Task 10: M25-DOGFOOD-09 Candidate Layout Radio Selection

Files:

  • Modify: apps/yune-web/source/src/Preferences.tsx

  • Modify if radio markup/styling needs adjustment: apps/yune-web/source/src/Inputs.tsx

  • Modify if labels need copy changes: apps/yune-web/source/src/consts.ts

  • Test: apps/yune-web/e2e/yune-typeduck.spec.ts

  • Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-09/

  • Step 1: Preserve the control baseline

    Keep candidate-layout-radio-report.json as the issue baseline. Current Preferences.tsx renders 候選排版 Candidate Menu Layout as Segment options backed by prefs.candidateMenuLayout; Inputs.tsx already exposes Radio.

  • Step 2: Add failing browser coverage

    Add a Playwright test that finds the Candidate Menu Layout field and asserts it contains radio inputs for horizontal/vertical instead of .yd-segment labels. Assert the accessible names include the existing labels from CANDIDATE_MENU_LAYOUT_LABELS.

  • Step 3: Replace the segmented control with radios

    In Preferences.tsx, replace the Segment mapping for candidateMenuLayout with Radio components using the same name, labels, prefs.candidateMenuLayout, and prefs.setCandidateMenuLayout. Do not change Segment globally because schema, Cangjie version, and other controls may still need segmented styling.

  • Step 4: Prove layout behavior remains intact

    In Playwright, select horizontal and vertical radio options and type a short input such as hai. Assert the selected radio state changes and the candidate panel uses the expected horizontal/vertical layout class or DOM shape for each option.

  • Step 5: Capture evidence

    Save desktop and narrow-viewport screenshots plus JSON under M25-DOGFOOD-09. Include radio labels, selected value before/after, and candidate panel layout evidence.

  • Step 6: Regenerate the TypeDuck-Web patch if source changed

    If any file under apps/yune-web/source/ changed, regenerate apps/yune-web/patches/yune-web-runtime.patch, reverse-check it from source/, and forward-check it on a clean source checkout.

Task 11: M25-DOGFOOD-10 IME Settings Horizontal Alignment

Files:

  • Modify: apps/yune-web/source/src/Preferences.tsx

  • Modify if shared layout styling is introduced: apps/yune-web/source/src/App.tsx

  • Modify if a reusable class is needed: apps/yune-web/source/src/index.css

  • Test: apps/yune-web/e2e/yune-typeduck.spec.ts

  • Evidence: apps/yune-web/e2e/results/m25-dogfooding/M25-DOGFOOD-10/

  • Step 1: Preserve the alignment baseline

    Keep ime-settings-margin-alignment-report.json as the issue baseline. Current App.tsx renders the playground stack directly inside main.m-auto.p-8.max-w-7xl, while Preferences.tsx adds a nested mx-auto max-w-6xl wrapper around 輸入法設定 IME Settings.

  • Step 2: Add failing visual-layout coverage

    Add a Playwright test that loads /web/ at a desktop viewport, locates the textarea or first stable playground content block above settings, locates the 輸入法設定 IME Settings wrapper, and compares bounding boxes. Assert Math.abs(playground.left - settings.left) <= 2 and Math.abs(playground.right - settings.right) <= 2.

  • Step 3: Add narrow-viewport coverage

    Repeat the bounding-box check at a narrow mobile viewport. Assert the settings wrapper does not introduce extra side margins compared with the textarea/main playground content and that no settings heading text overflows.

  • Step 4: Remove the nested width mismatch

    In Preferences.tsx, remove the nested mx-auto max-w-6xl constraint or replace it with a shared class that inherits the main content width. Prefer the minimal change: make the preferences root use the same available width as the content above, without changing main width, section card padding, or the two-column settings grid.

  • Step 5: Protect adjacent M25 layout changes

    Run the layout checks added for M25-DOGFOOD-06, M25-DOGFOOD-07, and M25-DOGFOOD-09 if they exist by implementation time. If they do not exist yet, capture a JSON note under M25-DOGFOOD-10 stating that the alignment fix preserved the current section order, checkbox/radio affordances, and settings grid class names.

  • Step 6: Capture evidence

    Save desktop and narrow-viewport screenshots plus JSON under M25-DOGFOOD-10. Include viewport sizes, left/right bounding-box values for the playground content and IME Settings wrapper, the pixel deltas, and the final preferences root class name.

  • Step 7: Regenerate the TypeDuck-Web patch if source changed

    If any file under apps/yune-web/source/ changed, regenerate apps/yune-web/patches/yune-web-runtime.patch, reverse-check it from source/, and forward-check it on a clean source checkout.

Execution Order After Intake

When the feedback list exists, execute in this order unless the ledger explicitly says otherwise:

  1. Reproduce and capture evidence for all Needs triage and Browser integration rows.
  2. Fix runtime/browser correctness rows before broad UI polish.
  3. Capture or reuse pinned oracle fixtures before any Engine correctness implementation.
  4. Batch adjacent UI polish only when it touches the same local components and does not blur issue ownership.
  5. Regenerate and reverse/forward check the TypeDuck-Web patch after every source-changing slice.
  6. Close ledger rows as evidence lands; do not wait until the end to update row status.
  7. Run focused gates for touched layers, then broad closeout gates if the batch changes shared behavior.

Closeout Gates

Before M25 can be archived:

cargo fmt
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace
npm.cmd --prefix packages/yune-typeduck-runtime test
npm.cmd --prefix packages/yune-typeduck-runtime run build
npm.cmd --prefix apps/yune-web/source run build
git diff --check

Run the real TypeDuck-Web Playwright tests for every closed browser-visible row. If source files under apps/yune-web/source/ changed, also run the patch reverse/forward checks from this plan.

Archive Rule

Archive this plan only after all M25 rows are Closed, Deferred, or Rejected with evidence/rationale. Update docs/roadmap.md and docs/requirements.md only for durable milestone status or new requirements, not for every intake row.