https://github.com/ydw99/regalia

Regalia — an open-source chess app. As of now, all the original codes are AI-generated.
https://github.com/ydw99/regalia

agpl-v3 ai-generated android-app chess multilanguage single-game stockfish-chess webview-app

Last synced: 7 days ago
JSON representation

Regalia — an open-source chess app. As of now, all the original codes are AI-generated.

• Host: GitHub
• URL: https://github.com/ydw99/regalia
• Owner: YDW99
• License: other
• Created: 2026-06-14T15:02:12.000Z (20 days ago)
• Default Branch: main
• Last Pushed: 2026-06-25T03:57:05.000Z (9 days ago)
• Last Synced: 2026-06-25T05:06:26.230Z (9 days ago)
• Topics: agpl-v3, ai-generated, android-app, chess, multilanguage, single-game, stockfish-chess, webview-app
• Language: HTML
• Homepage:
• Size: 3.93 MB
• Stars: 2
• Watchers: 2
• Forks: 0
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE&NOTICE.zip
  • Authors: AUTHORS-stockfish
  • Notice: NOTICE

Awesome Lists containing this project

README

          
# Regalia ♔

A standalone, open-source chess app for Android — play offline against Stockfish 18, analyze your games, and explore openings. No account, no network, no tracking. Now with **Chess960 (Fischer Random Chess)** support (v1.0.4).

"Regalia" is used solely as a project name for this open-source chess app. No trademark rights are claimed. Anyone is free to fork and rename their own version.

## Screenshots



  



*Portrait mode — evaluation bar, move history, AI opponent display with ponder info, and Control heatmap*

**Control Heatmap** — Tap the 🌗/🌈 button on the toolbar to toggle the control heatmap. Each square is dynamically colored by HSL to indicate which side controls it: blue-purple = your control, red = opponent's control, purple = contested. Hovering a square shows SVG arrows from each controlling piece to that square (warm gold for your pieces, cool silver-blue for opponent's). The info card below the board shows per-piece control contributions with position labels.

**🌿Line** — In the move record, 🌿 lines appear below each move showing engine analysis variations (MultiPV) and PGN import variations (RAV). Each variation is labeled 🌿Line 1, 🌿Line 2, etc., assigned sequentially by display order. PGN import variations are automatically parsed and displayed as 🌿Lines with proper move numbering. Toggle the Variations switch to show or hide them.

## Features

- **Stockfish 18 Engine** — arm64-v8a-dotprod variant with NEON acceleration for optimal performance on modern devices

- **Chess960 / Fischer Random Chess** (v1.0.4 NEW) — full support for the 960 starting positions with proper castling rules, SP-ID selector in New Game dialog, Shredder-FEN castling rights, and `UCI_Chess960` engine option

- **Standardized PGN** (v1.0.4 NEW) — import/export follows the 1994 PGN spec strictly: Seven-Tag Roster always emitted, `[%eval]` / `[%clk]` / `[%emt]` annotations embedded, Result terminator enforced, tolerant parser auto-corrects malformed input

- **NAG & Visual Annotations** (v1.0.4 NEW) — NAG ($1-$19) support; automatic selection & caching of `[%csl ...]` (square highlights) and `[%cal ...]` (arrows) per move: Square highlights — Blue=player net-control strong squares, Red=AI net-control strong squares, Yellow=high total-control squares, Green=neutral center squares; Arrows — Blue=multi-threat (one piece threatens 2+ enemy pieces), Red=check path, Yellow=queen-threat path, Green=escape squares

- **Time-Control Chess** (v1.0.4 NEW) — Sudden Death / Fischer Increment / Bronstein Delay / US Delay modes; live clock display with low-time warning; auto-emits `[TimeControl "..."]` header and `[%clk HH:MM:SS]` per-move annotations; for untimed games, emits `[%emt HH:MM:SS]` (elapsed move time)

- **Web Worker Pool** (v1.0.4 NEW) — `worker-pool.js` offloads PGN parsing, statistics computation, and control-map computation to a background thread; falls back to inline execution on devices without Worker support

- **8 Difficulty Levels** — from beginner (800 ELO) to maximum strength (2800+ ELO), plus Skill Level mode

- **PGN Import** — paste PGN from clipboard, or select a PGN file from your device

- **Review Mode** — full game replay with evaluation trend chart, move-by-move analysis, move classification (brilliant/good/blunder), and engine evaluation cache

- **MultiPV Analysis** — 1–8 lines of analysis simultaneously

- **ECO Opening Classification** — 500+ standard openings with search, category filtering, and book move recommendations

- **Syzygy Endgame Tablebases** — 7-piece endgame lookup via Lichess Tablebase API (requires network; auto-disables when offline)

- **Position Setup** — custom board editing with FEN copy/import

- **Ponder Mode** — engine thinks on opponent's time for stronger play

- **WDL Display** — Win/Draw/Loss probability shown alongside evaluation

- **Heatmap Control Statistics** (v1.0.4 NEW) — per-square average control across all positions, strongest/weakest square detection, center-control trend

- **Bilingual UI** — full Chinese/English toggle via the ↔️ button on the toolbar, with automatic system language detection

- **Landscape Support** — adaptive layout for both portrait and landscape orientation

- **Engine Configuration** — full UCI parameter control with export/import

- **Haptic Feedback** — responsive touch feedback throughout the interface

## Download

Download the latest APK from [GitHub Releases](https://github.com/YDW99/Regalia/releases). Enable "Install from unknown sources" to install.

## Requirements

- Android 5.0 (API 21) or later

- ARM64 device (arm64-v8a)

- ~80 MB storage

## Building

### Prerequisites

- Android SDK with API 35 (Android 15)

- Gradle 8.x (wrapper included)

- Stockfish 18 engine binary for arm64-v8a-dotprod

### Build Steps

1. Place the Stockfish engine binary at:

   ```

   /tmp/stockfish/stockfish-android-armv8-dotprod

   ```

2. Build the chess.html asset:

   ```bash

   cd src/main/assets/chess.src/

   bash build-chess.sh

   ```

3. Build the APK:

   ```bash

   ./gradlew assembleRelease

   ```

The signed APK will be at `build/outputs/apk/release/`.

## Project Structure

```

Regalia/

├── src/main/

│   ├── assets/

│   │   ├── chess.src/          # Source files (JS + CSS + HTML template)

│   │   │   ├── game-logic.js   # Chess rules, move generation, i18n

│   │   │   ├── chess960.js     # Chess960 SP-ID, Shredder-FEN, 960 castling rules (v1.0.4 NEW)

│   │   │   ├── pgn-standard.js # Standardized PGN encoder/decoder, NAG, [%csl]/[%cal], TimeControl (v1.0.4 NEW)

│   │   │   ├── worker-pool.js  # Web Worker pool for PGN/stats/control-map offloading (v1.0.4 NEW)

│   │   │   ├── ai-bridge.js    # Engine communication, eval display, PGN export

│   │   │   ├── ui.js           # Rendering, dialogs, interaction, review mode, time-control UI

│   │   │   ├── eco-data.js     # ECO opening classification data

│   │   │   ├── tablebase.js    # Lichess Syzygy tablebase queries + PGN import

│   │   │   ├── index.html.tpl  # CSS template

│   │   │   └── build-chess.sh  # Build script → chess.html

│   │   ├── chess.html          # Built output (combined JS+CSS+HTML)

│   │   ├── stats.html          # Statistics page (📊统计) — fullscreen WebView

│   │   ├── AGPLv3_Logo.svg     # AGPL logo for About page

│   │   └── GPLv3_Logo.svg      # GPL logo for 💾HTML export dialog

│   ├── java/com/Regalia/

│   │   ├── MainActivity.java   # WebView host, immersive mode, lifecycle

│   │   ├── StockfishNative.java # Engine process management, UCI protocol, SAF file I/O

│   │   │                       # v1.0.4: setChess960Mode() / isChess960Mode() bridge methods

│   │   ├── StatsActivity.java  # Fullscreen WebView for 📊统计 statistics page

│   │   ├── ChessWebViewClient.java # Page load handler

│   │   ├── EngineService.java  # Foreground service for engine stability

│   │   ├── ChessApp.java       # Application class, crash protection

│   │   ├── TlsSecurityHelper.java # TLS 1.2+ enforcement for tablebase API

│   │   └── RootDetector.java   # Informational root detection (About dialog)

│   ├── cpp/

│   │   ├── engine_jni.cpp      # JNI native chmod/renice (from DroidFish)

│   │   └── CMakeLists.txt

│   ├── res/

│   │   └── values/strings.xml

│   └── AndroidManifest.xml

├── NOTICE                      # Third-party component notices

├── NOTICE-DroidFish            # Original DroidFish notice

├── NOTICE-gradle               # Gradle notice (Apache v2.0)

├── AUTHORS-stockfish           # Stockfish project authors list

├── LICENSE-AGPL v3             # AGPL v3 full text (application)

├── LICENSE-GPL v3              # GPL v3 full text (engine components)

├── LICENSE-Apache v2.0         # Apache v2.0 full text (Gradle)

├── PRIVACY.md                  # Privacy policy

├── build-chess.py              # Python build script (alternative to build-chess.sh)

├── Regalia-v1.0.5-manual-zh.html  # Chinese user manual

├── Regalia-v1.0.5-manual-en.html  # English user manual

└── README.md

```

## Licensing

Regalia is a **combined work** under dual licensing:

| Component | License | File |

|-----------|---------|------|

| Original application code (UI, WebView, services, build scripts) | AGPL v3 | LICENSE-AGPL v3 |

| DroidFish-derived code (engine management, game logic, PGN parsing, UI patterns) | GPL v3 | LICENSE-GPL v3 |

| Stockfish 18 engine binary | GPL v3 | LICENSE-GPL v3 |

| ECO opening data | CC0 (data) / AGPL v3 (code) | — |

| Application icons | AI-generated / AGPL v3 | — |

Per GPL v3 Section 13, these licenses are compatible for combination. Each component retains its original license. Since AGPL v3 imposes stricter network interaction provisions (Section 13), its obligations effectively extend to the entire combined work, ensuring users who access the work over a network retain the right to obtain source code.

**Source code**: Available at https://github.com/YDW99/Regalia

### GPL v3 Files (DroidFish-derived)

- `StockfishNative.java` — Engine management logic

- `engine_jni.cpp` — Native chmod/renice from DroidFish

- `game-logic.js` — PGN disambiguation and SAN notation

- `ai-bridge.js` — Engine communication patterns

- `ui.js` — UI layout and interaction patterns

- `tablebase.js` — PGN parsing (GameTree/PgnToken/PgnScanner)

### AGPL v3 Files (original)

- `MainActivity.java`, `ChessApp.java`, `ChessWebViewClient.java`, `EngineService.java`

- `eco-data.js`, `index.html.tpl`, `CMakeLists.txt`

- `build-chess.py`, `build-chess.sh`

- All other project files not listed above

### Third-Party Components

- **DroidFish** — Engine management, game logic, PGN parsing, UI patterns (Copyright © Peter Österlund, GPL v3)

- **Stockfish 18** — Chess engine (Copyright © T. Romstad, M. Costalba, J. Kiiski, G. Linscott, GPL v3)

- **Lichess Tablebase API** — Endgame tablebase queries (public API, requires network)

- **lichess-org/chess-openings** — ECO opening classification data (CC0)

See [NOTICE](NOTICE) for full attribution details. More declaration documents are preserved in [NOTICE-DroidFish](NOTICE-DroidFish) and [AUTHORS-stockfish](AUTHORS-stockfish).

## Contributing

Contributions are welcome! Please ensure:

1. All contributions to the application layer are licensed under AGPL v3

2. Any modifications to DroidFish-derived or Stockfish code remain under GPL v3

3. Code is tested on physical Android devices (especially Xiaomi HyperOS 3)

## Version

**v1.0.5** (versionCode 105) — current release

The v1.0.5 release adds (Round-6 Revision 49):

1. **High aspect-ratio screen adaptation** — every interface scrolls vertically only,

   never horizontally, on any aspect ratio (ultra-tall phones, foldable inner screens,

   ultra-wide tablets).

2. **Notch/cutout/R-corner adaptation** — `shortEdges` cutout mode in AndroidManifest,

   `viewport-fit=cover` + `env(safe-area-inset-*)` CSS, auto-avoids notches/cutouts/R-corners.

3. **Sensor-fusion board anti-shake (OIS-style)** — new `StabilizationHelper.java` fuses

   `TYPE_LINEAR_ACCELERATION` (gravity-free translation) + `TYPE_GAME_ROTATION_VECTOR`

   (preferred, no magnetometer) / `TYPE_ROTATION_VECTOR` (fallback) sensors;

   board counter-shifts in real time to cancel device shake, plus ±2° roll-only

   tilt compensation. Toggle by long-pressing any board square (Toast in zh/en).

4. **Stats page "PGN Text" → "PGN Text (After processing)"** — clarifies the PGN is

   reconstructed/processed, not raw.

5. **Review arrow shrink** — arrows shrunk (4→2px stroke, 8×6→5×4 head) with per-arrow

   angular offset when multiple arrows share a target, so they stay distinguishable.

6. **Phase analysis precision** — rewritten per 国际象棋三阶段精确区分.md: multi-criteria

   detection (undeveloped minors, king castled, queen present, king advanced) replaces

   the old fixed material threshold.

7. **Manual chapter reorder** — Legal chapter moved to ch.2; manual changelog newest-first.

### Patch revisions under v1.0.5 (version number unchanged — still v1.0.5 / versionCode 105)

Multiple patch revisions have been applied under the same v1.0.5 version number.

The most significant recent revisions are:

- **Round-6 Revision 50** — Version-number residual cleanup (`v1.0.4` → `v1.0.5`

  in 3 UI locations); anti-shake Y-axis direction correction; PGN comment-merge

  fix (comments now attach to the PRECEDING move per PGN spec; only same-move

  comments merge with " | ").

- **Round-6 Revision 51** — App-launcher icon label synchronized to

  "Regalia v1.0.5"; anti-shake Y-axis direction further corrected.

- **Round-6 Revision 52** — Anti-shake screen-orientation auto-adapt (4 display

  rotations: 0°/90°/180°/270°); visual-annotation null-reference fix when

  `moveRecords[moveIdx]` is null (now uses `{csl:[],cal:[]}`).

- **Round-6 Revision 53** — Precise sensor re-selection: only

  `TYPE_LINEAR_ACCELERATION` + `TYPE_ROTATION_VECTOR` + `Display.getRotation()`

  (removed accelerometer/gyroscope/gravity/game-rot/geomagnetic-rot sensors).

  ±2° tilt compensation via `getRotationMatrixFromVector()`+`getOrientation()`.

  Yellow arrows extended to bidirectional queen threats (both sides' queens).

  Blue arrows also bidirectional. New `_computeInitialPositionAnnotations()`

  for the initial position (reviewStep 0).

- **Round-6 Revision 54** — Six improvements per user spec:

  1. **Yellow arrow extension** — when the mover's just-moved piece is a queen

     and its destination is attacked by an opp piece ("主动把皇后移动到受威胁

     格"), that yellow arrow's priority is BOOSTED to survive the cap-6 dedup.

  2. **Blue arrow extension** — when the mover's moved piece is itself at a

     threatened square ("主动把任意棋子移到受威胁格") AND this causes an opp

     piece to have 2+ threats on mover pieces, those blue arrows are BOOSTED.

  3. **Per-color arrow offsets** — each color has a FIXED diagonal offset

     (Blue→Upper-Left, Yellow→Lower-Left, Red→Upper-Right, Green→Lower-Right)

     applied to both start and end so overlapping arrows of different colors

     are visually separable.

  4. **Yellow square rounded-rectangle style** — yellow squares now render as

     a separate rounded-rectangle overlay (2px solid border, ~15% border-

     radius, ~10% inset) above the cell, leaving four-corner gaps so other

     colors' box-shadow insets remain visible underneath.

  5. **Roll-only board rotation** — `StabilizationHelper` rotation restricted

     to SCREEN-ROLL only (the steering-wheel tilt). Screen-pitch (nodding)

     and screen-yaw (door-like horizontal turn) are NOT compensated.

  6. **First-principles anti-shake audit** — cross-checked every sensor

     choice against the two uploaded sensor reference docs. Key change:

     PREFER `TYPE_GAME_ROTATION_VECTOR` (no magnetometer → no magnetic

     drift near laptops/metal desks) with `TYPE_ROTATION_VECTOR` fallback.

     Verified: `SENSOR_DELAY_GAME`, α=0.15 low-pass filter, 0.3° dead zone,

     decay-based drift prevention, lifecycle pattern, null-sensor fallback.

     Confirmed: `TYPE_ORIENTATION` / `device_orientation` is deprecated —

     we correctly use `Display.getRotation()`.

- **Round-6 Revision 55** — Line-by-line code audit of every source

  file, prioritized: bug fixes > features > perf > redundancy > simplification.

  1. **StatsActivity security parity** (bug) — the stats WebView was missing

     the defense-in-depth security flags that MainActivity has

     (`setAllowFileAccess(false)`, `setAllowContentAccess(false)`,

     `setAllowFileAccessFromFileURLs(false)`,

     `setAllowUniversalAccessFromFileURLs(false)`,

     `setMixedContentMode(NEVER_ALLOW)`, `setFilterTouchesWhenObscured(true)`).

     Added all of them so the stats page has the same security posture as the

     main game.

  2. **StatsActivity FLAG_FULLSCREEN fix** (bug) — was using the deprecated

     `FLAG_FULLSCREEN` unconditionally, which conflicts with Edge-to-Edge

     enforcement on Android 15+ (same black-screen bug MainActivity fixed in

     v18.4.6). Now gated on `SDK_INT < R`, matching MainActivity.

  3. **StatsActivity immersive mode** (feature) — the stats page now hides

     system bars (status + navigation) just like the main game, using the

     same platform-aware approach (`WindowInsetsController` on API 30+,

     legacy flags on API 21-29).

  4. **engineGoDepth eval-mode options** (bug) — `StockfishNative.engineGoDepth`

     (used for deep eval / analyze-all) was missing the

     `applyEvalModeOptions()` call that `engineEval` has. Without it, deep

     eval ran with gameplay `Contempt=24` (biased toward avoiding draws) and

     the user's `MultiPV` setting (potentially >1, reducing depth). Fixed by

     adding the call; the bestmove handler already restores gameplay settings.

  5. **getCtrlMap perf** (perf) — hoisted the attacker-position object out of

     the inner loop. Was allocating ~256 objects per ctrl-map (per render

     tick when heatmap is on); now allocates max 32 (one per piece) and

     reuses.

  6. **EngineService notification builder** (redundancy) — extracted shared

     `_buildNotificationWithContent()` helper used by both

     `buildNotification()` and `updateNotification()`, eliminating ~50 lines

     of duplicated builder configuration.

  7. **worker-pool fenToState consistency** — worker's check was

     `parts.length<4`, main thread's is `<2`. Changed worker to `<2` so

     minimal FENs (board + side-to-move only) are accepted by both paths.

  8. **chess960.js stale comment cleanup** — removed a misleading comment

     that claimed a line was "dead code" (it had been consolidated in

     v1.0.4). Code was correct; only the comment was stale.

- **Round-6 Revision 56** — Audit of `ai-bridge.js` MultiPV

  processing path and `ui.js` dirty-flag render pipeline. Six improvements:

  1. **MultiPV toggle-off stale lines** (bug) — `setConfigMultiPV(1)` mid-search

     left stale secondary PV lines (index 2..N) visible until the next

     `onBestMove` (seconds away). Now `_multiPVLines` and the display cache

     are cleared immediately, and the display refreshes instantly.

  2. **MultiPV display signature cache** (perf) — `_updateMultiPVDisplay()`

     was re-converting ALL PV lines from UCI to SAN on every progress tick

     (10-50×/sec), wasting 80+ `makeMv` calls per tick when the PV content

     was identical. Added `_multiPVDisplayCache` (per-line signature) to

     skip the conversion + DOM write when nothing changed.

  3. **MultiPV sort skip** (perf) — `onMultiPVProgress()` +

     `onEngineProgress()` sorted `_multiPVLines` unconditionally on every

     tick. Now only sort when a new line was appended (indices don't change

     on in-place update).

  4. **Eval display signature cache** (perf) —

     `_updateEvalDisplayIncremental()` rebuilt `innerHTML` on every

     `DIRTY_EVAL` tick even when eval values were unchanged (common during

     deep search). Added `_evalDispPrevSig` signature check to skip the DOM

     write; invalidated on full render.

  5. **MultiPV cache invalidation** (robustness) — wired

     `_clearMultiPVDisplayCache()` into all three MultiPV-reset paths

     (`onBestMove`, `restartCurrentEngine`, engine-recovery) so a recovered

     engine doesn't inherit stale display-cache signatures.

  6. **Dirty-flag pipeline audit** — confirmed `_performDirtyRender`'s

     flag-routing is correct; the TOOLBAR/PANEL/MOVES → full-render branches

     are technically dead code (always accompanied by bitcount>2) but left

     as-is for intent clarity and future-proofing.

- **Round-6 Revision 57** — CRITICAL FIX for blank-screen-on-launch

  regression introduced in Rev55. The Rev55 comment added to

  `worker-pool.js fenToState()` (inside the `_WORKER_SOURCE` template literal)

  used backticks around `parts.length<4`. Since `_WORKER_SOURCE` is a JS

  template literal (delimited by backticks), the backtick inside the comment

  **prematurely terminated** the template literal, causing a `SyntaxError`

  that crashed the entire `chess.html` bundle at parse time. Symptom: app

  launches to a blank screen showing only the background color. Fix: removed

  the backticks from the comment text. The actual code change from Rev55

  (`parts.length<4` → `parts.length<2`) is correct and unchanged. Verified

  with `node --check` on the bundled JS (exit code 0).

- **Round-6 Revision 58** — Two improvements per user spec:

  1. **Review-board arrow offset redesign** (ui.js) — Redesigned the arrow

     offset logic so: (a) each color's start offset = end offset (same bias

     direction, arrow direction preserved); (b) reverse-overlap (A→B and

     B→A) only happens for same-color arrows (they land on the exact same

     line; different-color reverse arrows are parallel but never collinear);

     (c) different-color arrows never overlap on the same line (guaranteed

     by four distinct diagonal biases: B→UL, Y→LL, R→UR, G→LR). Removed the

     Rev49 perpendicular "fan" offset that broke same-color reverse-arrow

     overlap. Added arrow deduplication by (color, from, to).

  2. **Magnetic-field-aware sensor switching** (StabilizationHelper.java) —

     On start, uses `TYPE_ROTATION_VECTOR` (accurate baseline via

     magnetometer). Background-monitors `TYPE_MAGNETIC_FIELD`; when magnetic

     disturbance detected (|B| deviation > 15 µT from rolling baseline),

     switches to `TYPE_GAME_ROTATION_VECTOR` (no magnetometer, immune to

     interference). When field calms (< 8 µT for 2 seconds), switches back

     to `TYPE_ROTATION_VECTOR`. Hysteresis prevents oscillation. Both sensors

     registered simultaneously (no re-registration gap).

- **Round-6 Revision 59** — CRITICAL FIX for rotation-axis bug

  in landscape orientations. The board's tilt-compensation rotation was

  controlled by the WRONG axis — it responded to screen pitch (front-back nod)

  instead of screen roll (steering-wheel tilt). Root cause:

  `SensorManager.getOrientation()` was called on the raw device-frame rotation

  matrix, which returns angles in the DEVICE's natural coordinate system, not

  the screen's. In landscape (ROTATION_90/270), device pitch and roll are

  swapped relative to the screen. Fix: Added

  `SensorManager.remapCoordinateSystem()` to transform the rotation matrix

  from device frame to screen frame BEFORE calling `getOrientation()`. After

  remapping, `orientationAngles[2]` is ALWAYS screen-roll (steering-wheel)

  regardless of display rotation. Simplified `dispatchTransform()` rotation

  to `-(smoothRoll - baselineRoll)` for ALL 4 orientations. Verified correct

  for 0°/90°/180°/270°. Also verified the Rev58 arrow offset logic already

  satisfies the user's requirements (same color = same bias for start+end;

  different colors = different biases).

- **Round-6 Revision 70** (current, 2026.6.27 net-control algorithm consistency fix) —

  Per user spec: the [%csl] blue/red net-control algorithm must match the

  "Square Control" info panel's net control. One fix:

  1. **Net-control consistency** (ui.js, BUG FIX) — blue/red candidates changed

     from one-sided dominance (`pAtk>0 && aAtk===0` / `aAtk>0 && pAtk===0`) to

     netCtrl-based (`netCtrl>0` / `netCtrl<0`), where `netCtrl = pAtk - aAtk`.

     This matches the info panel's `myCtrl - opCtrl`. A square with 3 white +

     1 black (netCtrl=+2) now correctly marks blue; previously it didn't (AI

     had an attacker, failing the one-sided condition). Score ranks by

     `|netCtrl|`. Both `_computeAndCacheVisualAnnotations()` and

     `_computeInitialPositionAnnotations()` updated. Spec comment items 3-4

     updated.

- **Round-6 Revision 69** (2026.6.27 multi-color per square + heatmap text deletion) —

  Per user spec: allow multi-color highlights on the same square; delete the

  heatmap conditional-display text from manuals/code. Two changes:

  1. **Multi-color per square** (ui.js, FEATURE) — `_rvCslMap` changed from

     `{square: {color}}` (single-color overwrite) to `{square: [colors]}`

     (array preserving all colors). Render code iterates colors: B/R/G accumulate

     into comma-separated box-shadow inset; Y emits yellow rounded-rect overlay.

     A square can now show both blue box-shadow AND yellow rounded-rect.

  2. **Heatmap text deletion** (manuals + stats.html, DOC) — deleted

     "不含热力图控制统计数据的 PGN → 不显示热力图控制统计区块" / English

     equivalent from v1.0.5 + v1.0.4 manuals (zh+en) and stats.html code comment.

- **Round-6 Revision 68** (2026.6.27 review yellow rounded-rect position fix) —

  Per user report: review-board yellow [%csl] rounded-rect overlay mispositioned.

  One fix + documentation audit:

  1. **Yellow rounded-rect position fix** (ui.js, CRITICAL BUG) — the yellow

     overlay div was emitted AFTER the cell div's `` instead of before,

     so `position:absolute` anchored to `.bgrid` (via `transform:translateZ(0)`)

     instead of the cell (`position:relative`). All yellow overlays stacked at

     the board origin. Fixed by moving the overlay emission to before the cell's

     `` so it becomes a child of the cell div. Parameters unchanged.

  2. **README/LICENSE/NOTICE audit** — checked all 14 matching files; added

     Rev68 entries to chess.src/README.license, java/README.license,

     Manual/README.license, NOTICE, README.md; added Rev54-68 summary entries

     to src/main/README.license, src/main/assets/README.license,

     src/main/res/README.license, src/main/cpp/README.license (were stalled at

     Rev53). LICENSE-* and NOTICE-* original texts need no update.

- **Round-6 Revision 67** (2026.6.27 landscape translation direction fix) —

  Per user request: rigorously verify translation anti-shake direction via first

  principles. Found ROTATION_90/270 boardPxX sign inverted (landscape horizontal

  shake amplified instead of cancelled). One fix:

  1. **Landscape boardPxX sign fix** (StabilizationHelper.java, CRITICAL BUG) —

     ROTATION_90: `boardPxX = -dispY` → `+dispY`; ROTATION_270:

     `boardPxX = +dispY` → `-dispY`. ROTATION_0/180 already correct. Derived

     per-orientation from first principles: device-natural Dx/Dy → world-axis

     mapping, then OIS (device world-right → board screen-left CSS -x; device

     world-up → board screen-down CSS +y). Comment rewritten as derivation

     table. Impact: landscape anti-shake horizontal direction now correct;

     portrait unaffected.

- **Round-6 Revision 66** (2026.6.27 yellow overlap + comment cleanup) —

  Per user spec: allow yellow squares to overlap with blue/red; comprehensively

  delete outdated/redundant comments. Two changes:

  1. **Yellow overlap allowed** (ui.js, FEATURE) — Rev65 EXCLUDED blue/red

     squares from yellow contention. Rev66 removes that exclusion — yellow now

     simply takes the top 3 by total attacker count, regardless of whether

     they're also blue or red. Rendering layer (box-shadow inset + yellow

     rounded-rect overlay) already supports multi-color highlights. Both

     `_computeAndCacheVisualAnnotations()` and

     `_computeInitialPositionAnnotations()` updated. Spec comment item 5 updated.

  2. **Comment cleanup** (10 files, REDUNDANCY) — audited and cleaned ~120

     lines of stale comments: StabilizationHelper.java Rev64 tombstones +

     Rev61 Y-axis fix; MainActivity.java wrong public-contract; ui.js

     arrow-rendering Rev49/54/58 narrative + Rev65/66 yellow explanations +

     Rev62 resign-sound + Rev63 arrow markers; index.html.tpl CSS multi-rev;

     worker-pool.js Rev57 backtick; chess960.js Rev62 array-swap + Rev55 meta;

     pgn-standard.js Rev62 lineLen; ai-bridge.js Rev62 _importedStartColor +

     mate:0/null + depth>0. Principle: retained license headers, current-

     behavior docs, algorithm explanations, i18n, pitfall warnings; deleted

     tombstones describing deleted code / fixed bugs / multi-rev narratives.

- **Round-6 Revision 65** (2026.6.27 yellow-square logic fix) —

  Per user spec: yellow squares (non-arrow) should be the squares with the

  highest total control count (total attackers, regardless of piece color).

  One change:

  1. **Yellow-square logic fix** (ui.js, BUG FIX) — Rev61's condition

     `pAtk>0 && aAtk>0` (both sides must have attackers) was too strict,

     excluding high-total one-sided squares. Rev65 changes it to `total>=2`

     (all squares with 2+ total attackers are candidates), then excludes

     squares already chosen as blue/red (one-sided dominance), and picks the

     top 3 by total. This correctly implements "黄色格子=双方总控制数量最多的

     那几个格子". Both `_computeAndCacheVisualAnnotations()` and

     `_computeInitialPositionAnnotations()` updated. Spec comment item 5 updated.

- **Round-6 Revision 64** (2026.6.27 remove rotation + undo cache sync) —

  Per user spec: remove the rotation part of anti-shake (keep only translation),

  delete all rotation code/comments; and after undo, revert the corresponding

  content in the background real-time cached PGN. Two changes:

  1. **Anti-shake rotation removed** (StabilizationHelper.java + index.html.tpl,

     FEATURE) — `StabilizationHelper.java` completely refactored. Deleted

     sensors: `TYPE_ROTATION_VECTOR`, `TYPE_GAME_ROTATION_VECTOR`,

     `TYPE_MAGNETIC_FIELD` (only `TYPE_LINEAR_ACCELERATION` retained). Deleted

     methods: `processRotationVector()`, `processMagneticField()`. Deleted all

     rotation state fields, constants, scratch arrays. CSS `.bwrap.stabilized`

     `transform` changed from `translate3d(...) rotate(var(--stab-rot,0deg))` to

     `translate3d(var(--stab-x,0px),var(--stab-y,0px),0)` only. `applyTransform()`

     no longer sets `--stab-rot`; added `removeProperty('--stab-rot')` to clear

     residual values. File slimmed from 720 → 337 lines (53% reduction).

     Benefits: 4 sensors → 1 sensor, lower battery, eliminates the entire class

     of rotation-direction bugs from Rev59-63.

  2. **Undo reverts cached PGN content** (ui.js, BUG FIX) — `undoMove()` now

     calls new `_invalidateCachesForUndoneMoves(currentMoveCount)` after

     restoring `moveRecords`. Clears three caches: `_reviewEvalCache` (key >

     currentMoveCount), `_visualAnnotationsCache` (key >= currentMoveCount,

     '_initial' sentinel kept), `_cachedOriginalPGN` (set to null). This

     ensures stale eval/annotation data for undone moves doesn't leak into

     re-played moves, review mode, or PGN export. `_redoStack` is NOT cleared

     (undo must remain reversible).

- **Round-6 Revision 63** (2026.6.27 anti-shake + arrow optimization) —

  Per user spec: limit anti-shake rotation to ±10°, optimize arrow

  positioning and redesign arrow shape. Three changes:

  1. **Anti-shake rotation limit ±10°** (StabilizationHelper.java, FEATURE) —

     `MAX_ROTATION_DEG` reduced from ±30° (Rev61) to ±10°. First-principles:

     sensor floor ~1°, human perception ~0.5-1°, handheld tilt rarely >8°,

     1:1 mapping preserved in 1°-10° range. Dead zone 1.0° unchanged.

  2. **Arrow offset optimization** (ui.js, FEATURE) — offset magnitude changed

     from `Math.min(3, cellSize*0.08)` (3px cap + 8%) to `cellSize*0.12` (12%

     no cap). Scales proportionally with cell size so offset fraction is

     constant across screens. Verified safe for cellSize 20-60px (max radial

     offset = cellSize/2 - 2.75, we use 12% which is well within bound).

     Cell-center formula `col*cellSize + cellSize/2` confirmed exact for all

     cell sizes.

  3. **Arrow shape redesign** (ui.js, VISUAL) — minimal arrow: thin 1.5px

     line (was 2px) + compact 4×3 triangular arrowhead (was 5×4) + butt

     linecap + no origin dot. Line shorten 9→4px (matches new arrowhead).

     stroke-opacity 0.95→0.9 (pieces underneath remain readable). The minimal

     representation that still communicates origin, target, direction, type.

- **Round-6 Revision 62** (2026.6.27 second-pass audit) — Second-pass

  first-principles code audit. One critical bug fix + 5 bug fixes + 3 perf

  improvements + redundancy cleanup:

  1. **Chess960 SP-ID bishop-color array swap fix** (chess960.js, CRITICAL BUG)

     — `_CH960_LIGHT_BISHOP_FILES` and `_CH960_DARK_BISHOP_FILES` were swapped.

     a1/c1/e1/g1 are DARK squares (not light), b1/d1/f1/h1 are LIGHT (not dark).

     This broke 720/960 SP-IDs; SP-ID 518 produced "RNQBBKNR" instead of

     "RNBQKBNR". After fix, all 960 SP-IDs pass round-trip test. The parity

     checks in `backRankToSPID()` were fixed in sync.

  2. **Tablebase 404 misidentified as server-down** (tablebase.js, BUG) — 404

     means "position not in tablebase", not "server down". 3 × 404 falsely set

     `_tbOffline=true` for 60s. Now only 5xx + network errors count.

  3. **PGN move-number regex missing 'O'** (worker-pool.js + tablebase.js, BUG)

     — `(?=[a-hKQRBN])` didn't match 'O', so "1 O-O" wasn't normalized to

     "1. O-O" and "1" was mis-tokenized. Fixed to `[a-hKQRBNO]`.

  4. **Resign sound never played** (ui.js, BUG) — `_resignGame()` called

     undefined `_playSound('lose')`. Now calls `playSound('gameover')`.

  5. **mate:0 vs mate:null distinction lost** (ai-bridge.js, BUG) — `c.mate||0`

     lost the checkmate-now distinction. Fixed to `c.mate!=null?c.mate:0`.

  6. **worker-pool Map delete-during-iteration** (worker-pool.js, BUG) —

     `for...of` + `.delete()` risked skipping entries on some engines. Fixed

     to snapshot keys via `Array.from()` first.

  7. **ECO recommendation LRU no-refresh-on-hit** (game-logic.js, PERF) — cache

     hit returned without delete+re-insert, so hot entries could be evicted

     before cold ones. Fixed to refresh LRU on hit (matching `_tbCache` pattern).

  8. **MultiPV display wasted work during AI thinking** (ai-bridge.js, PERF) —

     computed hintParts/signatures/cache checks 10-50×/sec during AI thinking

     but discarded the result. Early-exit when `!isHintLoading`.

  9. **Eval chart label loop LRU churn** (ui.js, PERF) — used `get()` (refreshes

     LRU) instead of `peek()` in a read-only iteration. Fixed to `peek()`.

  10. **Redundancy cleanup** — removed dead `_importedStartColor` ternary

      (variable never declared, guard always true); removed dead `let lineLen=0`

      in composePGN; simplified redundant `depth>0` check; simplified two

      unreachable ternary branches.

  11. **chess960.js comment correction** — `randomSPID()` comment claimed

      "P(>8 retries) ≈ 0.4%"; actual is (256/65536)^8 ≈ 1.3e-19. Code correct,

      comment fixed.

  12. **Known Chess960 castling detection limitation** (DOCUMENTED, NOT FIXED)

      — `Math.abs(to.col-from.col)===2` only detects castling when king starts

      on col 4. In Chess960 (king can start on cols 1-6), ~78.8% of positions

      have broken castling detection. HIGH risk to fix (4 coordinated locations

      + Zobrist hash); deferred to a dedicated task. Standard chess unaffected.

- **Round-6 Revision 61** (2026.6.27) — Per the 2026.6.27

  development plan. Two critical bug fixes + one feature refinement +

  four additional bug fixes from a first-principles code audit:

  1. **Board anti-shake Y-axis direction reversal fix** (StabilizationHelper.java,

     CRITICAL BUG) — all 4 screen-orientation cases (ROTATION_0/90/180/270)

     in `dispatchTransform()` had the WRONG sign on `boardPxY`. OIS requires

     that when the device moves UP, the board moves DOWN (CSS +Y) to

     compensate; the original code moved the board UP — amplifying the shake.

     User report: "屏幕突然向上移动，棋盘不但不向下位移来抵消，反而也向上".

     X-axis signs were already correct.

  2. **Board rotation sensitivity adjustment** (StabilizationHelper.java,

     FEATURE COMPLETION) — `ROTATION_DEADZONE_DEG` 0.3° → 1.0° (filter

     sensor noise + sub-degree hand tremor); `MAX_ROTATION_DEG` 45° → 30°

     (covers normal handheld tilt, avoids visual abruptness). 1:1 ratio

     (|board rotation| = |screen roll delta|) strictly preserved within

     1°–30° range. User report: "棋盘旋转角度过于敏感，屏幕小幅度转动，

     棋盘就大幅度转动".

  3. **Review-board yellow-square logic fix** (ui.js, CRITICAL BUG) —

     yellow-square selection changed from `total>=2` to `pAtk>0 && aAtk>0`

     (both sides must have attackers). Previously, single-side-dominated

     squares (e.g. 5 white + 0 black = total 5) were misclassified as

     yellow — they should be BLUE. User spec: "黄色格=双方总控制高格(当前

     未正确实现)".

  4. **Arrow stroke-linecap changed to butt** (ui.js, VISUAL REFINEMENT) —

     changed from "round" to "butt" in review-board SVG arrows. The round

     linecap created a small semi-circle at the arrow's start that read as

     an "extra dot" — contradicting user spec "箭头末端没有多余的圆点".

  5. **Engine extraction progress percentage fix** (StockfishNative.java,

     BUG FIX) — `total / 114115752L * 13` truncated to 0 for total < 114 MB

     (Java left-to-right evaluation). Fixed to `total * 13L / 114115752L`.

  6. **stats.html unclosed-brace infinite-loop fix** (stats.html, CRITICAL

     BUG) — `while(moveText.includes('{'))...` would ANR on malformed PGN

     with unclosed `{`. Capped to 10 iterations.

  7. **StatsActivity stream leak fix** (StatsActivity.java, BUG FIX) —

     `loadAssetAsBase64()` switched to try-with-resources to guarantee

     InputStream closure on exception paths.

- **Round-6 Revision 60** — Three changes per user spec:

  1. **Arrow origin dot removed** (ui.js) — removed the small filled circle

     at each arrow's origin in the review-board SVG. The arrowhead suffices.

  2. **Arrow offset calibrated** (ui.js) — reduced per-color diagonal offset

     from 14%/6px-max to 8%/3px-max to prevent arrows from exceeding cell

     boundaries on small screens.

  3. **Rotation limit expanded** (StabilizationHelper.java) — increased

     from ±2° to ±45°. The board can now counter-rotate

     up to 45° to compensate for large device tilt.

     (Note: Rev61 later reduced this back to ±30° — see Rev61 entry above.)

**v1.0.4** (versionCode 104) — previous major release

The v1.0.4 release adds:

1. **Chess960 (Fischer Random Chess) mode** — full implementation with SP-ID selector,

   Shredder-FEN castling rights, 960 castling rules, `UCI_Chess960` engine option,

   and `[Variant "Chess960"]` PGN tag.

2. **Standardized PGN import/export** — Seven-Tag Roster always emitted, `[%eval]`

   / `[%clk]` annotations embedded, Result terminator enforced, tolerant parser

   auto-corrects malformed input.

3. **Heatmap-based control statistics** — per-square average control across all

   positions, strongest/weakest square detection, center-control trend.

4. **Version sync** — versionCode, versionName, AndroidManifest, build.gradle,

   every in-app UI display, Java file VERSION constants, README, NOTICE, every

   README.license file, and HTML manuals all synchronized to v1.0.4.

### Patch revisions (version number unchanged — still v1.0.4 / versionCode 104)

Multiple patch revisions have been applied under the same v1.0.4 version number.

The most significant recent revisions are:

- **Round-5 Revision 17** — Xiaomi HyperOS 3 cache-clearing fix: added a

  Java-side persistent storage layer (`AndroidBridge.persistentGet/Set/Remove`)

  backed by `SharedPreferences`, dual-written alongside WebView localStorage so

  the review-eval cache, language preference, and crash-recovery state survive

  HyperOS 3's aggressive cache wipes. Also added `AndroidManifest.xml` backup

  rules (`backup_rules.xml` + `data_extraction_rules.xml`) declaring

  SharedPreferences and WebView Local Storage as user data (not cache).

- **Round-5 Revision 18** — 📚 PGN Cache Manager: new modal dialog

  accessible from the review toolbar (📚 button next to 🌗/🌈). Lets users

  save the current PGN to a named entry, list all saved entries (with size +

  mtime), import any entry with one tap (syncs both review and main move

  list), and select multiple entries for batch deletion. Backed by

  `AndroidBridge.listPGNCaches/savePGNCache/getPGNCache/deletePGNCaches`

  writing to `/data/data/com.Regalia/files/pgn_cache/` (HyperOS-proof).

  Also removed the 200-entry cap on `_reviewEvalCache` — now unlimited.

- **Round-5 Revision 19** — onclick syntax fix + click-empty-moves-to-review:

  fixed HTML attribute-escaping bug that broke the cache manager's click

  handlers. Also made the "No moves yet" text in the main move list

  clickable to enter review mode (useful for users who just installed the

  app and want to import a PGN before playing any move).

- **Round-5 Revision 20** (current) — HyperOS 3 eval-cache root-cause fix +

  PGN cache rename/tags + review toolbar reorder:

  • **Eval cache**: first-principles analysis revealed that

    SharedPreferences.apply() is async — HyperOS 3 SIGKILLs the app before

    the disk flush, losing data. Fix: dedicated file

    (`/data/data/com.Regalia/files/eval_cache.json`) with atomic write

    (tmp + fsync + rename) + synchronous `saveEvalCacheSync()` Java

    interface. Critical-event flush on onPause/onStop/onUserLeaveHint +

    pagehide/visibilitychange/beforeunload. Debounce reduced 500ms→150ms.

    Eval cache now loads synchronously at JS module construction — **instant

    recovery** of all previously-analyzed evals when entering review mode.

  • **PGN Cache Manager**: new ✏️ Rename and 🔖 Add Tags buttons per entry.

    Tags stored in `.tags.json`, displayed as gold capsule chips.

    Max 10 tags per entry, each ≤30 chars, case-insensitive dedup.

    listPGNCaches() returns all tags in one call.

  • **Review toolbar button order**: 🌗/🌈 moved to right of 💬Vars toggle,

    left of 📝 PGN — groups display buttons (Vars, Heatmap) before action

    buttons (PGN, Save, Cache, FEN).

  • New Java interfaces: persistentSetSync, persistentFlush,

    loadEvalCacheSync, saveEvalCacheSync, renamePGNCache, setPGNCacheTags,

    getPGNCacheTags.

  • backup_rules.xml + data_extraction_rules.xml: added eval_cache.json

    and pgn_cache/ as user data for cloud backup / device migration.

  • Dead code cleanup: removed unused `_pgnCacheToast` variable.

- **Round-5 Revision 21** — Eval cache writes now synchronous +

  PGN cache tag filter/search:

  • **Eval cache**: `_reviewEvalCache.set()` / `.delete()` / `.clear()` now

    trigger **immediate synchronous write** to `eval_cache.json` (no

    debounce). Each eval arrives ~1/sec via `onEngineEval`; sync write

    overhead is ~1-5ms — imperceptible. Completely eliminates the 150ms

    debounce window where new eval data could be lost if HyperOS 3

    SIGKILLed the app mid-debounce.

  • **PGN Cache Manager**: new 🔍 text search box + tag-chip quick-filter

    row at the top of the modal. Search matches entry names AND tags

    (case-insensitive substring). Tag chips collect all unique tags across

    all entries; click a chip to filter by that tag. Tag chips below each

    entry name are also clickable. Smart Select All/None operates only on

    visible (filtered) entries. Filter status bar shows match count.

    Auto-clears filter on close/reopen.

  • New JS functions: `_pgnCacheOnSearchInput`, `_pgnCacheApplyFilter`,

    `_pgnCacheClearFilter`, `_pgnCacheFilterByTag`,

    `_pgnCacheEntryMatchesFilter`, `_pgnCacheCollectAllTags`.

  • New i18n keys: `pgn_cache_search_placeholder`, `pgn_cache_search_apply`,

    `pgn_cache_search_clear`, `pgn_cache_filter_all`, `pgn_cache_filter_by_tag`,

    `pgn_cache_filter_status`, `pgn_cache_filter_no_match` (all zh/en).

- **Round-5 Revision 22** — New Game dialog layout + HyperOS 3

  defense strengthened + PGN cache search UX:

  • **New Game dialog**: Chess960 toggle moved to be directly below the

    "Play Color" section (was below AI Book Moves). Removed the redundant

    "Classic Openings (Optional)" residual title from the Chess960-OFF

    branch. The Chess960-ON dedicated-settings section heading is now

    `T('chess960_label')` (was `T('classic_openings')`).

  • **HyperOS 3 eval-cache defense**: `MainActivity.onDestroy()` now also

    calls `_flushReviewEvalCache()` + `persistentFlush()` before destroying

    the WebView. Previously onDestroy skipped the flush — any eval data

    that arrived between onStop and onDestroy could be lost. The four-layer

    defense is now: (1) dedicated file `eval_cache.json` with atomic write;

    (2) synchronous `saveEvalCacheSync()` using `FileDescriptor.sync()`;

    (3) critical-event flush on Java onPause/onStop/onUserLeaveHint/

    onDestroy AND JS pagehide/visibilitychange/beforeunload; (4) Rev21's

    per-set synchronous write (no debounce).

  • **PGN Cache Manager search**: removed the redundant 🔍 apply button

    (soft-keyboard Enter now triggers search directly via

    `enterkeyhint="search"`). Input type changed to `search` for native

    affordances. The ✏️/🔖/📥 buttons per entry now share the same row

    (was a vertical column).

  • **HTML manuals**: removed the obsolete "Stockfish engine binary"

    paragraph from the v1.0.3 changelog in both zh and en manuals.

- **Round-5 Revision 23** — First-principles code review fixes:

  • **Critical bug fix**: Engine restart (`onEngineRestarting()` /

    `restartCurrentEngine()`) no longer calls `_reviewEvalCache.clear()`,

    which previously destroyed the ENTIRE persisted eval cache (all games'

    data). Now only resets `_reviewEvalRequestedStep=-1` — Stockfish is

    deterministic, so other games' evals remain valid.

  • **Performance**: Eval cache writes restored to 150ms debounce (Rev21's

    per-set synchronous write serialized the entire ~12MB Map on every eval

    callback). `_flushSync()` from lifecycle handlers still forces immediate

    write, so the debounce window is covered by Rev22's onDestroy flush.

  • **Bug fix**: `saveEvalCacheSync()` now uses `Files.move()` with

    `ATOMIC_MOVE` + `REPLACE_EXISTING` on API 26+ — the old delete+rename

    sequence had a brief window where neither file existed.

  • **Redundancy**: Removed dead second cache check in `requestEngineEval()`,

    dead `moveNum` variable in `_buildPGNString()`, stale 🔍 comment.

- **Round-5 Revision 24** (current) — Analyze-all root-cause fix + instant

  PGN-comment eval recovery + player rename + PGN cache completeness:

  • **Analyze-all interruption root-cause fix**: First-principles analysis

    revealed FIVE distinct causes of "分析全部 interrupted at a certain

    point":

    (1) The old `startStep=0` fallback when all steps were cached caused

    `requestEngineEval()` to early-return on cache hit WITHOUT calling

    `_reviewAnalyzeAdvance()`, stalling the batch until the safety timer

    fired. Now `reviewAnalyzeAll()` detects the all-cached case up front

    and short-circuits to completion INSTANTLY (no engine calls, no timer).

    (2) The safety timer was set ONCE with `min(N*8s, 300s)` — capped at 5

    minutes. For long games with depth-22 evals (5-15s/step), the 5min cap

    fired mid-analysis. Now the per-step safety timer RESETS on every

    advance (60s/step, no global cap), and a stuck step is SKIPPED (not

    aborting the whole batch).

    (3) If the engine auto-recovered (`onEngineReady` after Java

    `recoverEngine()`), the batch was silently dropped. Now `onEngineReady`

    resumes the batch if `_reviewAnalyzeAllActive` is still true.

    (4) If `requestEngineEval()` hit cache during batch (rare race or

    re-entry), it returned without advancing. Now it detects

    `_reviewAnalyzeAllActive + cache hit` and calls `_reviewAnalyzeAdvance`.

    (5) `onEngineError` no longer clears `_reviewAnalyzeAllActive` — the

    per-step safety timer covers the recovery window, and the batch resumes

    on `onEngineReady`.

  • **Instant cache restoration from PGN comments**: `tablebase.js _parsePGN()`

    now extracts `[%eval ...]` tags from `{}` comments BEFORE stripping them,

    and `importPGN()` populates `_reviewEvalCache` from the extracted evals.

    When the user enters review mode after import, all positions with

    `[%eval ...]` annotations are shown INSTANTLY with their cached eval —

    the engine is only invoked for positions WITHOUT `[%eval ...]`. This

    implements the "instant recovery from cache (including PGN comments)"

    requirement. Supports all Lichess eval formats: `[%eval 0.35]`,

    `[%eval -1.5]`, `[%eval #5]`, `[%eval #-3]`, `[%eval M5]`.

  • **PGN cache archive completeness**: `_pgnCacheSaveCurrent()` now ALWAYS

    uses `_buildPGNString(true)` (forceIncludeVariations=true). The previous

    behavior preferred `_cachedOriginalPGN` (the imported text), which lost

    NEW moves played after import. `_buildPGNString()` now includes: all

    moveRecords (including post-import moves), variations (forced on

    independent of showVariations UI toggle), `[%eval ...]` from the review

    cache, `[%emt]/[%clk]/[%csl]/[%cal]`, Seven-Tag Roster + supplementary

    tags. The saved PGN archive contains the COMPLETE game record.

  • **Player rename feature**: Clicking the "你"/"You" name on the main

    interface player bar opens a rename prompt. The new name is:

    - Persisted via `AndroidBridge.persistentSet('Regalia_humanName', ...)`

      (HyperOS 3-proof)

    - Reflected in all PGN text (`[White "..."]` / `[Black "..."]`)

    - Reflected in PGN cache archives and clipboard copies

    - Loaded synchronously at JS module init (available before first render)

    Entering an empty string or the default name resets to "你"/"You".

    PGN import with explicit `[White "..."]`/`[Black "..."]` headers also

    populates `_humanPlayerName` if the human player's slot has a non-default

    name.

  • **Redundant 📥 button removed**: The 📥 Import button in the PGN Cache

    Manager was redundant — clicking the entry's name already imports it.

    Removed to save horizontal space; ✏️ Rename and 🔖 Tags buttons remain.

  • **New i18n keys**: `rename_player_hint`, `rename_player_prompt`,

    `rename_player_saved`, `rename_player_reset` (all zh/en).

  • **New globals**: `_humanPlayerName`, `playerWhite`, `playerBlack`.

  • **Stockfish 18 arm64-v8a-dotprod**: Engine binary updated to the

    official sf_18 release (NDK r27c build, 114 MB, NEON+dotprod

    acceleration).

- **Round-5 Revision 25** (current) — Stats page promotion symbols + Chess960

  dialog heading cleanup + first-principles review:

  • **Stats page promotion symbol fix (bug fix)**: In `stats.html

    renderPGNText()`, the `renderSAN()` function previously only prepended

    the pawn symbol for pawn moves (e.g., "e8=Q" → "♙e8=Q") but left the

    "=Q"/"=R"/"=B"/"=N" promotion suffix as a literal letter — inconsistent

    with non-pawn moves where the piece letter IS converted to a symbol.

    Now every promotion piece (Q/R/B/N after '=') is converted to its

    side-colored symbol (♕/♛, ♖/♜, ♗/♝, ♘/♞), regardless of whether the

    move is a pawn move or a piece move. Examples: "e8=Q+" → "♙e8=♕+"

    (White) / "♟e8=♛+" (Black); "exd8=R" → "♙exd8=♖" (White) / "♟exd8=♜"

    (Black). A shared `_replacePromo()` helper handles both the pawn-move

    and piece-move branches uniformly, including the rare theoretical case

    of a piece-letter move with a promotion suffix (not legal in standard

    chess, but defensive).

  • **Chess960 dialog heading removed (redundancy cleanup)**: In the New

    Game dialog, when Chess960 is enabled, the dedicated settings section

    (SP-ID input, back-rank preview, note) previously had a redundant

    `
` heading showing "菲舍尔任意制象棋" / "Fischer Random Chess". The

    Chess960 toggle above (with its own label) already makes it clear what

    mode the user is in. Rev25 removes the heading so the settings appear

    directly under the toggle, saving vertical space. The heading is

    removed in both Chinese and English modes.

- **Round-5 Revision 26** — Stats page promotion symbol root-cause

  fix:

  • **Root cause of "e8=Q still shows letter" found and fixed**: Rev25 added

    a `_replacePromo()` helper to convert `=Q`/`=R`/`=B`/`=N` suffixes to

    piece symbols, but the promotion STILL showed as a letter because the

    underlying SAN regex in `stats.html renderPGNText()` never captured the

    promotion suffix for non-capture pawn promotions. The regex's pawn

    non-capture alternative was `[a-h][1-8][+#]?` (NO promotion group), so

    for "e8=Q" it matched only "e8" — leaving "=Q" to be rendered as raw

    literal characters via the "pass through other characters" fallback.

    Rev26 fixes the regex alternative to `[a-h][1-8](?:=[QRBN])?[+#]?`,

    which captures the full "e8=Q" so `renderSAN()` (and its `_replacePromo`

    helper) can convert the promotion piece to a symbol. The same fix is

    applied to the variation-move regex. Now ALL promotion formats render

    correctly: `e8=Q` → ♙e8=♕, `exd8=R` → ♙exd8=♖, `a8=N#` → ♟a8=♞#,

    `b1=B+` → ♙b1=♗+. Verified via regex test against 12 sample moves.

- **Round-5 Revision 27** — Hyperlinks open in system browser +

  Resign feature + Stats board display + multi-report P0/P1 fixes:

  • **Clickable hyperlinks (About page + everywhere)**: The About page's

    "Source code: https://github.com/YDW99/Regalia" is now a real ``

    hyperlink (previously plain text). ALL hyperlinks anywhere in the app

    (About dialog's AGPL/GPL license links, stats.html links, etc.) are

    intercepted by a new global click handler that calls

    `AndroidBridge.openUrlInBrowser(url)` — a new `@JavascriptInterface`

    method in StockfishNative.java and StatsActivity.java that launches

    the system default browser via `Intent.ACTION_VIEW`. Defense-in-depth:

    `ChessWebViewClient.shouldOverrideUrlLoading()` and the stats WebView's

    WebViewClient also intercept http(s) URLs at the navigation layer, so

    links triggered by any means (JS `window.open`, meta refresh, etc.)

    are routed to the system browser. URLs are strictly validated to be

    http(s) only — file:, content:, javascript:, intent: schemes are

    silently rejected.

  • **🏳️ Resign feature (DeepSeek review 2.1)**: A new 🏳️ Resign button

    appears in the player bar next to the "Your turn" indicator. Tapping

    it opens a confirmation dialog; confirming ends the game with the

    opponent winning. PGN export follows the 元宝 PGN report convention:

    `[Result "0-1"]` if White resigns / `[Result "1-0"]` if Black resigns,

    `[Termination "Resignation"]` supplementary tag, and a

    `{White resigns.}` / `{Black resigns.}` comment on the last move.

    Stops the engine (`stop` command + isAIThinking=false) and the game

    clock. The game-over overlay shows the correct resign text and

    win/lose emoji.

  • **Stats page board display (DeepSeek review)**: A small chess board

    now appears ABOVE the PGN text panel in 📊统计, showing the position

    after the currently-selected move (or initial position if nothing is

    selected). The board uses the SAME color scheme as the main

    chess.html board (SQ_LIGHT/SQ_DARK gradients, coordinate labels with

    stroke, piece symbols with color/stroke/glow), with White at the

    bottom and Black at the top. Width is responsive (max 360px, aspect

    ratio 1:1).

  • **Kimi stats.html P0 fix — Backspace global interceptor**: The

    `keydown` listener that turns Backspace into "return to main game"

    now checks whether the key event originated from an INPUT, TEXTAREA,

    or contentEditable element. Previously, pressing Backspace while

    editing the PGN paste field or any prompt input would (1) prevent

    the character deletion AND (2) immediately bounce the user back to

    the main game — making text editing impossible.

  • **Kimi stats.html P0 fix — applySANMove king-safety validation**:

    `applySANMove` now validates that the moving side's king is NOT in

    check AFTER the candidate move, by calling a new `_isKingSafe()`

    helper. Previously, `canMoveTo` only checked pseudo-legal moves, so

    a pinned piece could "move" and leave its own king in check —

    corrupting the replayed position and all downstream statistics.

    Castling also now checks post-castle king safety. If a disambiguated

    piece fails the check, the scanner continues to the next matching

    piece (correct SAN disambiguation behavior).

  • **Kimi stats.html P1 fix — game complexity threshold**: Changed

    `validEvals.length>=3` to `>=2`. With 2 evals, `changes` has 1

    element — `avgChange` and `variance` are both well-defined

    (variance=0 for a single sample). The previous >=3 check skipped

    the entire complexity section for 2-eval games.

  • **Kimi stats.html P1 fix — _escFEN escaping**: Added `"` and `'`

    escaping to `_escFEN()`, matching chess.html's `_esc()` behavior.

    Closes a potential XSS vector where a malicious FEN imported via

    PGN could break out of HTML attribute context.

  • **Kimi stats.html P1 fix — exportFullHTML regex**: Replaced the

    lazy `[\s\S]*?\n}` regex (which could match a nested object

    literal's closing brace instead of the function's) with a new

    `_stripFnBody()` helper that does brace-depth counting to find the

    function's true closing brace. Prevents orphaned code from leaking

    into the exported HTML.

  • **Stats.html hyperlink interceptor**: New global click handler in

    stats.html routes `` clicks to the new

    `AndroidBridge.openUrlInBrowser()` bridge method.

  • **formatEval resign fix**: The resign case in `formatEval()` now

    uses `_resignWinnerColor` directly (set by `_resignGame()`) instead

    of trying to detect "White wins" / "Black wins" in the gameOver

    text (which only contains "resigns"). Fixes a bug where the player

    always appeared to lose even when they won (e.g. AI resigns).

- **Round-5 Revision 28** (current) — Stats↔main PGN sync + resign emoji

  fix + stats board promotion fix:

  • **Stats↔main PGN sync**: Every time the user enters the stats page

    (📊 button), `openStatsPage()` now ALWAYS rebuilds the PGN from the

    current `moveRecords` via `_buildPGNString(true)` (forceIncludeVariations=true).

    Previously, it used `_cachedOriginalPGN` (the text from the last

    importPGN/importPGNFile call), which became STALE when the user played

    new moves after importing — the stats page would show the imported

    PGN's moves, not the current game's moves. Now the stats page always

    reflects the current main/review game state.

  • **Stats→main PGN import-back prompt**: When the user imports a PGN on

    the stats page (via 🗃️ Paste PGN or 📂 Select PGN File) and then

    returns to the main activity (back button or 返回对局 button), a new

    "🗃️ Import PGN to game?" dialog appears with Yes/No/Cancel:

    - **Cancel**: Stay on the stats page (don't return to main).

    - **No**: Return to main without importing (main keeps its current game).

    - **Yes**: Return to main AND trigger the existing "💾 Save PGN file?"

      prompt (to avoid losing the main/review's current PGN), then import

      the stats-page PGN via `importPGN()`.

    The dialog UI matches the app's existing dialog style (.dov/.dlg/.dlg-btns).

    The Android back button is intercepted: if the import-back dialog is

    visible, back = Cancel (dismiss); otherwise back = returnToGame()

    (which may show the import-back dialog if a PGN was imported).

  • **Resign game-over emoji fix**: The game-over overlay's emoji was

    hardcoded to show 🤝 for all non-checkmate terminations. Now it shows

    🏳️ when `_gameOverStatusKey==='resign'` (matching the resign button's

    icon), and 🤝 only for genuine draws (stalemate, 50-move, repetition,

    insufficient material, agreement).

  • **Stats board promotion fix**: The stats page board display (added in

    Rev27) showed a pawn on the back rank instead of the promoted piece

    (queen/rook/bishop/knight) for promotion moves. Root cause: the

    `moveRecords.push({...})` in `tablebase.js _parsePGN()` (the PGN import

    path) was missing the `promotion` field, so `openStatsPage()` serialized

    move records without promotion info, and the stats page's

    `executeMove()` couldn't apply the promotion. Fixed by adding

    `promotion: move.promotion||null` to the pushed move record.

- **Round-5 Revision 29** — PGN import failure root-cause fix + stats page

  en passant + comprehensive bug fixes:

  • **CRITICAL: PGN import failure fix**: Rev28's `promotion:

    move.promotion||null` referenced an undefined variable `move` (loop

    variable is `parsedMove`). The ReferenceError was caught by the

    surrounding try/catch and shown to the user as "Invalid PGN format" —

    EVERY PGN import failed. Fixed to `(parsedMove&&parsedMove.move&&

    parsedMove.move.promotion)||null`.

  • **CRITICAL: Stats page en passant support**: The simplified SAN parser's

    `canMoveTo()` and `executeMove()` did not handle en passant. Importing

    a PGN with en passant (e.g. move 48 `gxf6` in the user's

    Regalia_game 3Ω.pgn) caused parsing to fail at move 94 → side-to-move

    desync → 25 subsequent moves all rejected → stats page showed only 98

    moves. Added en passant branch to `canMoveTo()` and captured-pawn

    removal to `executeMove()`. Stats page now parses all 123 moves.

  • **Stats page checkmate symbol fix**: `buildSAN()` previously never

    emitted `#` (checkmate) — all checks shown as `+`. Added full

    checkmate detection via in-place simulation (mutate→check king

    safety→revert) to avoid the recursive executeMove→buildSAN→

    executeMove loop that caused stack overflow with the naive approach.

  • **NAG Black-side fix**: `ai-bridge.js _buildPGNString()` and

    `stats.html classifyMove()` used raw White-POV delta for mate

    thresholds. A BLACK mating move was misclassified as $4 (blunder)

    instead of $3 (brilliant). Fixed to use side-relative moverDelta.

  • **Stats page Black-to-move-start caption fix**: `_renderStatsBoard()`

    caption used `lastIdx%2===0` for White/Black, ignoring

    `parsed.isBlackToMoveStart`. For black-to-move-start games, the color

    was inverted. Fixed to respect `parsed.isBlackToMoveStart`.

  • **PGN parsing tolerance improvement**: Unclosed `{...}` comments

    previously survived the multi-iteration regex removal, and the

    residual content was tokenized into bogus move tokens generating

    meaningless NULL placeholders. Now follows PGN spec's tolerant

    behavior: an unclosed `{` consumes the rest of the movetext.

  • **Variation prefix number falsy-0 fix**: `v.prefixEllipsisNum||vmn`

    in `_buildPGNString()` incorrectly fell back to vmn when

    prefixEllipsisNum===0 (legitimate rare value). Fixed to explicit

    null check.

  • **stats.html PGN parse error i18n**: Was hardcoded English "PGN parse

    error". Added i18n key `pgn_parse_error` (zh: "PGN 解析失败").

  • **PGN import high-tolerance**: Invalid moves are skipped with NULL

    placeholders (shown as dimmed "—") and side-to-move is auto-advanced

    to avoid cascade failure. Nested variations supported via recursive

    descent parser. App self-exported PGN re-imports perfectly (123

    moves → export → re-import → 123 moves, lossless).

- **Round-5 Revision 30** (current) — Robustness hardening + scroll

  restoration bug fix:

  • **ROBUSTNESS: Scroll position restoration fix**: The .mlist move list

    repeatedly jumped back to top during re-renders. Root cause: with

    `scroll-behavior:smooth` on .mlist, programmatic `scrollTop`

    assignment during restore fired intermediate 'scroll' events that

    overwrote `_mlistScrollState.scrollTop`, causing the NEXT render to

    restore to a stale intermediate position. Fixed with a

    `_scrollRestoreGuard` flag that suppresses `_onMlistScroll` during

    programmatic restore, plus temporarily switching `scroll-behavior`

    to 'auto' for instant restoration. The save phase now also reads

    `scrollTop` DIRECTLY from the live DOM (not just from the event

    handler) for the most reliable snapshot. Same pattern applied to

    `.review-body` in review mode. `_scrollRestoreGuard` is reset on

    new game.

  • **ROBUSTNESS: ChessWebViewClient URL overload**: Now also overrides

    the newer `WebResourceRequest`-based `shouldOverrideUrlLoading`

    overload (API 24+). The deprecated String-based overload only fires

    on API 21-23; on API 24+ only the `WebResourceRequest` overload is

    invoked. Without this override, http(s) links clicked on API 24+

    devices would NOT be redirected to the system browser. The new

    overload delegates to the String overload for consistent behavior.

  • **ROBUSTNESS: chess960.js randomSPID() modulo bias elimination**:

    Previously `buf[0]%960` had a slight bias (65536 isn't a multiple of

    960; values 0..255 were ~0.03% more likely than 256..959). Now uses

    rejection sampling: only accept values < 65280 (the largest multiple

    of 960 ≤ 65536) before applying %960. Bounded retry count (8)

    prevents infinite loops; fallback to %960 if all retries exceed

    (extremely unlikely — P ≈ 0.4%).


See `NOTICE` (VERSION HISTORY SUMMARY → v1.0.4) and the per-module

`README.license` files for full details.

- **Round-5 Revision 31** (current) — Stats board piece sizing + review

  orientation scroll fix + UCI timed-ponder compliance + dead code cleanup:

  • **Stats board piece sizing fix**: `stats.html _renderStatsBoard()` piece

    font-size was `5.2vw` (viewport-relative). In landscape, the viewport is

    wide (e.g. 800px+) but the board container is capped at 360px, so pieces

    overflowed their squares. Fixed: piece font-size now uses a CSS variable

    `--piece-size` set via JS to `(boardWidth / 8 * 0.72)`. This scales

    correctly with the ACTUAL board container in both portrait and landscape,

    on all Android WebView versions (no container query support needed).

    Added `_updateStatsBoardPieceSize()` called after render and on

    resize/orientationchange.

  • **Review mode orientation-change scroll bug fix**: After orientation

    change, the review move list jumped to a wrong scroll position because

    the save phase read `scrollTop` from the OLD layout and the restore

    phase applied that pixel offset to the NEW layout (where it points to a

    different move). The review-moves-list scroll-into-view logic didn't

    fire either (no `reviewStep` change). Fixed: (1) on orientationchange,

    force `_lastReviewStepScrolled=-2` so the next render re-centers the

    active move in the NEW layout; (2) invalidate `_mlistScrollState.valid`

    so the main move list re-snapshots from the new DOM; (3) set

    `_skipReviewBodyScrollRestore` flag so the review-body doesn't restore

    the stale scrollTop (the active-move scroll-into-view handles

    re-centering instead).

  • **UCI COMPLIANCE: Timed-game ponder fix**: Per UCI spec, `go ponder`

    MUST include time params (wtime/btime/winc/binc) so the engine can

    switch to normal time management when `ponderhit` fires. Previously

    `startPonder()` sent only "go ponder" with no time params — breaking

    timed-game ponder (engine searched indefinitely or panicked after

    ponderhit). Fixed: `startPonder()` now accepts 4 time params; JS side

    passes current `gameClocks` values. For untimed games, all zeros are

    passed and Java side sends plain "go ponder". Added `@deprecated`

    legacy single-arg `startPonder()` overload for backward compat.

  • **Dead code cleanup**: Removed unused `_unclosed` flag in

    `tablebase.js _parsePGN` (was set but never read). Removed dead

    `oppKing` variable in `stats.html buildSAN()` (was always null, never

    read — leftover from abandoned check-detection). Removed dead

    `ch==='\s'` comparison in `pgn-standard.js` (a single char can never

    equal the 2-char string "\s" — typo for regex \s class). Fixed

    `composePGN()` to avoid leading blank lines when tagPairs is empty.

- **Round-5 Revision 32** (current) — ECO book move engine-unresponsive

  bug fix + UCI eval optimization:

  • **CRITICAL: ECO book move causing engine unresponsive**: When "AI

    prefers ECO opening book moves" is enabled, after 3 consecutive

    book-served AI moves, the engine appeared "unresponsive" — no AI

    move was made. Root cause: `doAIMove()` increments `_aiRetryCount`

    on every call (expecting it to be reset by `onBestMove`). When the

    ECO book provides a move, the engine is NOT called, so `onBestMove`

    never fires, and `_aiRetryCount` accumulates. After 3 book moves,

    `_aiRetryCount>=3` → `doAIMove()` falsely concludes "AI move failed

    after 3 consecutive timeouts" → shows `ai_timeout` toast and

    RETURNS WITHOUT CALLING THE ENGINE. Fix: reset `_aiRetryCount=0`

    on a successful book move (and tablebase move) — the "request" was

    satisfied, just by the book instead of the engine.

  • **UCI EVAL OPTIMIZATION (per SF18 best-practices doc)**:

    `engineEval()` and `engineEvalDeep()` now set `Contempt=0` (objective

    eval; default 24 biases scores by avoiding draws, distorting

    analysis), `MultiPV=1` (max search depth; MultiPV>1 reduces depth),

    `UCI_ShowWDL=true` (Win/Draw/Loss probability output) BEFORE the

    eval search. After the eval, `handleBestMove()`'s STATE_EVAL case

    calls `restoreGameplayOptions()` to restore `Contempt=24` (aggressive

    gameplay) and the user's MultiPV setting. New methods:

    `applyEvalModeOptions()`, `restoreGameplayOptions()`.

- **Round-5 Revision 33** (current) — Seldepth (SD) display throughout:

  • **Seldepth parsing & display**: Per SF18 eval best-practices doc, added

    seldepth (selective search depth / tactical depth) parsing and display.

    Seldepth reflects the actual max depth reached in tactical variations

    (usually >= depth). Displayed as "SD" right after "D" to match

    the existing abbreviated depth style (e.g., "D15 SD22").

  • **Main UI eval bar + AI opponent bar**: Both now show "D15 SD22" during

    engine search. SD only shown when > 0 AND > depth (seldepth == depth

    is redundant).

  • **Review mode eval bar**: Now shows real-time depth (D), seldepth (SD),

    nodes, and nps during analysis — previously only showed D with no

    nodes/nps. This gives the user full visibility into the engine's

    search progress while reviewing a position.

  • **PGN eval annotation**: formatEvalTag() now includes "SD" in the

    `{SF18 D15 SD22 +0.5 50%W/30%D/20%L}` annotation when seldepth > depth.

  • **Implementation**: StockfishNative.java new SELDEPTH_PATTERN +

    processInfoLine() parsing + 9th/6th/7th params to onEngineProgress/

    onPonderProgress/onEngineEval. ai-bridge.js new _sfSeldepth global +

    cache storage. ui.js updated all 3 eval bar render paths.

- **Round-5 Revision 34** (current) — Chess960 castling own-piece capture

  fix + heatmap coordinate labels:

  • **CRITICAL: Chess960 castling could capture own pieces**:

    `isChess960CastlingLegal()` only checked squares BETWEEN the king and

    rook, not the king's/rook's destinations. In Chess960, the king and

    rook can start close together (e.g., king b1, rook c1), so the king's

    path extends far beyond the rook through occupied squares. `makeMv()`

    would overwrite those squares, destroying own pieces. Fix: now checks

    the UNION of the king's path and rook's path — every square must be

    empty except the king's and rook's own starting positions.

  • **stats.html Chess960 castling**: `applySANMove()` no longer hardcodes

    king from col 4 (e1) — scans the back rank for the king's actual

    column. `executeMove()` castling rook move is now Chess960-aware,

    finding the nearest rook on the castling side instead of always

    using col 7/0.

  • **stats.html heatmap coordinate labels**: Each heatmap cell now

    displays its coordinate (e.g., "a8", "e4") at the top, centered.

- **Round-5 Revision 35** (current) — Timed-game engine stop fix + AI bar

  seldepth label localization:

  • **CRITICAL: Timed-game engine search exceeded clock time**: Two root

    causes fixed:

    (1) Clock expiry didn't send "stop" to the engine. The engine kept

        searching past the GUI's 0-second mark. Fix: `_onGameClockExpired()`

        now calls `AndroidBridge.engineStop()` for an immediate hard-stop.

    (2) `wtime`/`btime` sent to the engine were stale (measured at JS call

        time, but "go" is sent up to 6s later after stopAndWait + ucinewgame

        + readyok). The engine over-allocated search time. Fix: `engineGoTimed()`

        now deducts the setup overhead + a safety margin from the clock

        values before sending "go".

  • **New @JavascriptInterface methods**: `engineStop()` (hard-stop any

    engine state, discards bestmove), `sendToEngine(String)` (raw UCI

    command fallback).

  • **AI opponent bar SD label localization**: The AI opponent bar now

    shows "选深" (Chinese) / "SelDepth" (English) instead of "SD". The

    eval bar keeps "SD" for compactness. New i18n key: `seldepth_label`.

- **Round-5 Revision 36** (current) — AI bar layout restructure + seldepth

  colon + critical bug fixes:

  • **AI opponent bar layout restructure**: Engine search real-time info

    (depth/seldepth/nodes/nps) moved to line 2 right-aligned. Ponder info

    moved to line 3 right-aligned. Line 1 shows only compact status.

  • **Seldepth label colon**: Added ':' between seldepth label and value

    (选深:22 / SelDepth:22) for format consistency.

  • **CRITICAL: engineStop() idle-state fix**: No longer sets

    _discardingPonderBestmove when engine is idle — prevents the next

    game's first bestmove from being silently discarded.

  • **CRITICAL: onPonderProgress score perspective fix**: was inverted

    (playerColor==='black' → should be 'white'). Ponder scores now display

    with correct sign.

  • **stats.html Chess960 castling detection**: Fixed to use destination

    column (6/2) instead of distance (===2) for Chess960 compatibility.

  • **Redundancy**: _updateAIThinkDisplay() no longer duplicates the

    "thinking" placeholder on line 2.

- **Round-5 Revision 37** (current) — AI bar captured pieces split at 7:

  • **AI opponent bar captured pieces split**: When the AI's captured pieces

    count exceeds 7, pieces 8+ are displayed on line 3 (left-aligned),

    sharing the line with ponder info (right-aligned). Pieces 1-7 remain

    on their original row. The player bar keeps the original single-row

    wrap behavior (unchanged). `capturedPiecesHtml()` gains a `splitAt`

    parameter (7 for AI bar, 0/omitted for player bar).

- **Round-5 Revision 38** (current) — AI opponent bar overflow fix:

  • **AI opponent bar overflow fix**: Some information was exceeding the AI

    bar's display width. Fixed by adding `overflow:hidden` to the inner

    column div, `max-width:100%` + `box-sizing:border-box` to

    `#ai-search-info`, `min-width:0` + `overflow:hidden` to the line-3

    container, `min-width:0` to `#ai-ponder-info`, and `max-width:100%` +

    `overflow:hidden` to `.pbar` CSS. All text now properly truncates with

    ellipsis when the bar is too narrow.

- **Round-5 Revision 39** (current) — Ponder info not displayed when captured

  pieces > 7:

  • **CRITICAL: Ponder info not displayed when captured pieces > 7**: The

    `#ai-cap-overflow` div had `width:100%` + `flex:0 0 auto`, causing it to

    take ALL available width in the line-3 flex container, pushing

    `#ai-ponder-info` off-screen. Fix: removed `width:100%`, changed to

    `flex:0 1 auto` (shrink-to-fit) + `min-width:0`. Now the overflow pieces

    only take the space they need, leaving room for ponder info on the right.

- **Round-5 Revision 40** (current) — Overflow pieces horizontal fix +

  ponder alignment fix:

  • **Overflow pieces vertical wrapping fix**: `#ai-cap-overflow` used

    `flex-wrap:wrap` causing pieces to stack vertically when container

    narrowed. Changed to `flex-wrap:nowrap` + `overflow:hidden` — pieces

    always stay horizontal. Reduced font-size to 0.85rem for overflow pieces.

  • **Ponder alignment fix**: Line-3 container used

    `justify-content:space-between` which put `#ai-ponder-info` at LEFT

    when no overflow (single child). Changed to `margin-left:auto` on

    `#ai-ponder-info` for always-right alignment.

- **Round-5 Revision 41** (current) — Player bar captured pieces font-size

  + stats board coordinate labels rework:

  • **Player bar captured pieces font-size**: Reduced from 1.1rem to 0.85rem

    to match AI opponent bar (both bars now visually consistent).

  • **Stats page board coordinate labels rework**: File labels (a-h) moved

    outside board on TOP, rank labels (1-8) moved outside board on LEFT,

    per-cell coordinate labels (e.g. "e4") added inside each cell — all

    matching main board style exactly (color, stroke, glow, paint-order).

- **Round-5 Revision 42** (current) — AI bar overflow pieces not displayed:

  • **CRITICAL: Overflow captured pieces not displayed**: `#ai-cap-overflow` had

    `overflow:hidden` (clipped pieces) + `flex:0 1 auto` (shrank to 0 width).

    Fix: removed `overflow:hidden`, changed to `flex:0 0 auto` (natural width,

    no shrink). All overflow pieces now display correctly.

- **Round-5 Revision 43** (current) — Exported stats HTML stuck at "Loading..." fix:

  • **CRITICAL: Exported stats HTML stuck at "Loading..."**: The `_stripFnBody()`

    function's brace-counting logic matched nested function closing braces

    (at depth 0 + next char \n/space), truncating the strip of `_exportFullHTML`

    mid-function. This left orphaned code in the exported JS → syntax error →

    page never executed → stuck at "Loading...". Fix: changed end condition to

    depth===0 AND '}' at start of line (preceded by \n) — standard JS convention

    for top-level function closing braces. All 9 stripped functions now strip

    correctly; full exported JS passes syntax validation.

- **Round-5 Revision 44** (current) — Eval bar stuck fix + Worker memory leak

  fix + _stripFnBody simplification:

  • **Eval bar stuck mid-analysis**: Added JS-side eval safety timer (10s) that

    resets `_evalLoading` if the engine doesn't respond. Prevents the eval bar

    from getting stuck at "分析中" when `stopAndWaitForBestmove` blocks or the

    engine fails to respond.

  • **Worker memory leak fix (review report item 1)**: Added `terminateWorker()`

    registered on `pagehide`/`beforeunload` — terminates the Web Worker and

    revokes its Blob URL on page exit, preventing OOM from accumulated Worker

    threads in Android WebView.

  • **_stripFnBody simplification (review report item 18)**: Replaced fragile

    brace-counting with simple regex `function name(){...}` → `function name(){}`

    via non-greedy match to next `\n}` at column 0. Verified for all 9 functions.

- **Round-5 Revision 45** (current) — Engine timeout threshold extended:

  • **AI safety timer**: Extended from 15s to 30s to accommodate timed-game

    mode where the engine may need 20-25s for complex positions at high

    difficulty levels.

  • **Eval safety timer**: Extended from 10s to 15s to match (eval searches

    at depth 15/22 can take 10-12s on complex positions).

- **Round-5 Revision 46** (current) — Timeout emoji fix + engine timeout 360s:

  • **Timeout emoji fix**: Game-over overlay and eval bar now show ⌛ for timeout

    wins (was 🤝 — fell through to default draw emoji).

  • **Engine timeout extended**: AI safety timer 30s→360s (6 min); eval safety

    timer 15s→30s. Fully accommodates long timed games.

- **Round-5 Revision 47** (current) — Timeout game-over text re-localization:

  • **Timeout text not re-localizing**: The timeout game-over text was set

    directly using `_lang` instead of `T()`, so switching language after

    game-over didn't update the text. Fix: added `'timeout'` branch to

    `_gameOverStrFromStatus()` with new `_timeoutWinnerColor` variable and

    new i18n key `timeout_win_suffix`. Now re-localizes correctly.

- **Round-5 Revision 48** (current) — PGN comment completeness + Blue/Yellow

  arrows + review board visual annotations + stats conditional display:

  • **PGN comment/visual-annotation completeness fix**: First-principles

    review revealed that `_parsePGN` in `tablebase.js` STRIPPED all brace

    comments during parsing — only `[%eval ...]` tags were extracted

    beforehand. This meant `[%csl ...]` (square highlights), `[%cal ...]`

    (arrows), AND free-text comments from imported PGNs were LOST. The user

    requirement "all valid () variations and {} comments must be fully

    received" was not met. Fix: added a pre-strip extraction pass that

    walks the movetext character-by-character (tracking paren depth for

    variation-internal comments), extracts `[%csl]`/`[%cal]` tags AND

    free-text comment bodies, and attaches them to the next main-line

    move. `importPGN()` now populates `_visualAnnotationsCache` and

    `mr.comment` from the extracted data. `_buildPGNString()` now includes

    `mr.comment` in the comment parts (with literal `{`/`}` escaped to

    Unicode full-width braces to avoid premature comment termination).

    PGN round-trip is now lossless for comments and visual annotations.

  • **NEW: Blue and Yellow arrows in visual annotations**: Per user spec

    (备忘1.md), added two new arrow colors:

    - Blue arrow = one side's piece simultaneously threatens multiple

      (>1) enemy pieces (arrows from threatening piece to each

      threatened piece)

    - Yellow arrow = one side's threat to the other side's queen

      (arrow from threatening piece to queen's square)

    `_computeAndCacheVisualAnnotations()` in `ui.js` now computes these

    from the post-move control map. Caps: top 3 attackers by threat

    count (blue), top 3 attackers by piece value (yellow), max 4 blue

    arrows per attacker.

  • **Stats page visual annotations section enhancement**: The visual

    annotations section in `stats.html` now displays all 4 arrow colors

    (Blue, Red, Yellow, Green) in the SAME left-to-right order as the 4

    square colors — so same colors align VERTICALLY (per user spec).

    New i18n keys: `blue_arrows_threats`, `yellow_arrows_queen` (zh+en).

    Updated `visual_annotations_desc` to mention all 4 arrow types.

  • **Stats page conditional display**: Per user spec, the visual

    annotations section is now ONLY shown if the PGN actually contains

    `[%csl]` or `[%cal]` tags. Similarly, the heatmap control statistics

    section heading is now inside the `_posCount>0` check.

  • **Review board visual annotations overlay**: Per user spec (备忘1.md),

    when the control heatmap is OFF in review mode, the review board now

    displays `[%csl]`/`[%cal]` annotations:

    - Square highlights via CSS box-shadow inset (3px, matching last-move

      hint width). Colors: B=#4a90d9, R=#e74c3c, Y=#f1c40f, G=#27ae60.

    - Arrows drawn in an SVG overlay layer (same style as the main board's

      `_updateArrows()`: 4px stroke, round linecap, 0.85 opacity,

      triangular arrowhead marker per color).

    The `.review-board` CSS rule now has `position:relative` so the SVG

    overlay's `position:absolute` anchors correctly.