## Summary
Adds Xteink X3 hardware support to CrossPoint Reader. The X3 uses the
same SSD1677 e-ink controller as the X4 but with a different panel
(792x528 vs 800x480), different button layout, and an I2C fuel gauge
(BQ27220) instead of ADC-based battery reading.
All X3-specific behavior is gated by runtime device detection — X4
behavior is unchanged.
Depends on community-sdk X3 support: open-x4-epaper/community-sdk#19
(merged).
## Changes
### HAL Layer
**HalGPIO** (`lib/hal/HalGPIO.cpp/.h`)
- I2C-based device fingerprinting at boot: probes for BQ27220 fuel
gauge, DS3231 RTC, and QMI8658 IMU to distinguish X3 from X4
- Detection result cached in NVS for fast subsequent boots
- Exposes `deviceIsX3()` / `deviceIsX4()` helpers used throughout the
codebase
- X3 button mapping (7 GPIOs vs X4's layout)
- USB connection detection and wake classification for X3
**HalDisplay** (`lib/hal/HalDisplay.cpp/.h`)
- Calls `einkDisplay.setDisplayX3()` before init when X3 is detected
- Requests display resync after power button / flash wake events
- Runtime display dimension accessors (`getDisplayWidth()`,
`getDisplayHeight()`, `getBufferSize()`)
- Exposed as global `display` instance for use by image converters
**HalPowerManager** (`lib/hal/HalPowerManager.cpp/.h`)
- X3 battery reading via I2C fuel gauge (BQ27220 at 0x55, SOC register)
- X3 power button uses GPIO hold for deep sleep
### Display & Rendering
**GfxRenderer** (`lib/GfxRenderer/GfxRenderer.cpp/.h`)
- Buffer size and display dimensions are now runtime values (not
compile-time constants) to support both panel sizes
- X3 anti-aliasing tuning: only the darker grayscale level is applied to
avoid washed-out text on the X3 panel. X4 retains both levels via
`deviceIsX4()` gate
**Image Converters** (`lib/JpegToBmpConverter`, `lib/PngToBmpConverter`)
- Cover image prescale target uses runtime display dimensions from HAL
instead of hardcoded 800x480
### UI Themes
**BaseTheme / LyraTheme** (`src/components/themes/`)
- X3 button position mapping for the different physical layout
- Adjusted UI element positioning for 792x528 viewport
### Boot & Init
**main.cpp**
- X3 hardware detection logging
- Adjusted init sequence for X3 (no `HalSystem::begin()` dependency on
X3 path)
**HomeActivity**
- Uses runtime `renderer.getBufferSize()` instead of static
`GfxRenderer::getBufferSize()`
FYI I did not add support for the gyro page turner. That can be it's own
PR.
## Summary
Replace per-pixel getRenderMode() + rotateCoordinates() + bounds checks
with a DirectPixelWriter struct that pre-computes orientation and render
mode state once per row. Use bitwise ops instead of division/modulo for
cache pixel packing. Skip PNG cache allocation when buffer exceeds 48KB
(framebuffer size) since PNG decode is fast enough that caching provides
minimal benefit, and the large buffer competes with the 44KB PNG decoder
for heap.
## Additional Context
Measured improvements on ESP32-C3 @ 160MHz:
- JPEG decode: 5-7% faster (1:1 scale)
- PNG decode: 15-20% faster (1:1 scale)
- Cache renders: 3-6% faster across both formats
- Eliminates "Failed to allocate cache buffer" errors for large PNGs
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**< PARTIALLY >**_
## Summary
- Add CSS `display: none` support to the EPUB rendering pipeline (fixes
#1431)
- Parse `display` property in stylesheets and inline styles, with full
cascade resolution (element, class, element.class, inline)
- Skip hidden elements and all their descendants in
`ChapterHtmlSlimParser`
- Separate display:none check for `<img>` tags (image code path is
independent of the general element handler)
- Flush pending text blocks before placing images to fix layout ordering
(text preceding an image now correctly renders above it)
- Bump CSS cache version to 4 to invalidate stale caches
- Add test EPUB (`test_display_none.epub`) covering class selectors,
element selectors, combined selectors, inline styles, nested hidden
content, hidden images, style priority/override, and realistic use cases
## Summary
* **What is the goal of this PR?** Fixing / extending the hyphenation
logic to deal with words containing an apostophe as raised in #1186
* **What changes are included?**
## Additional Context
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**PARTIALLY**_ (as the
user provided a thorough analysis that I followed)
## Purpose
This PR includes some preparatory changes that are needed for an
upcoming performant CJK font feature. The changes have no impact on
render time and heap allocation for latin text. **Despite this, I think
these changes stand on their own as a better font
compression/decompression implementation.**
## Summary
- Font decompressor rewrite: Replaced the 4-slot LRU group cache with a
two-tier system — a page buffer (glyphs prewarmed before rendering
begins) and a hot-group fallback (last decompressed group retained for
non-prewarmed
glyphs).
- Byte-aligned compressed bitmap format: Glyph bitmaps within compressed
groups are now stored row-padded rather than tightly packed before
DEFLATE compression, improving compression ratios by making identical
pixel rows produce
identical byte patterns. Glyphs are compacted back to packed format on
demand at render time. Reduces flash size by 155 KB.
- Page prewarm system: Added `Page::collectText` and
`Page::getDominantStyle` to extract per-style glyph requirements before
rendering, and `GfxRenderer::prewarmFontCache` to pre-decompress only
the groups needed for the dominant style
— eliminating mid-render decompression for the common case.
- UTF-8 robustness fixes: `utf8NextCodepoint` now validates continuation
bytes and returns a replacement glyph on malformed input;
`ChapterHtmlSlimParser` correctly preserves incomplete multi-byte
sequences across word-buffer flush
boundaries rather than splitting them.
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**YES**_ Architecture and
design was done by me, refined a bit by Claude. Code mostly by Claude,
but not entirely.
## Summary
**What is the goal of this PR?**
### Problem
Inter-word gap widths were computed as two separately-snapped integers:
```cpp
gap = getSpaceWidth(fontId, style); // fp4::toPixel(spaceAdvance)
gap += getSpaceKernAdjust(fontId, leftCp, rightCp, style); // fp4::toPixel(kern1 + kern2)
```
Because `fp4::toPixel(a) + fp4::toPixel(b)` can differ from
`fp4::toPixel(a + b)` by +/-1 pixel when the fractional parts straddle a
rounding boundary, each inter-word space could be one pixel wider or
narrower than the correct value. This affected line-break width
decisions and word-position accumulation across the whole paragraph
layout pipeline.
### Fix
Replaces `getSpaceKernAdjust()` with `getSpaceAdvance(fontId, leftCp,
rightCp, style)`, which combines the space glyph advance and both
flanking kern values (`kern(leftCp, ' ')` + `kern(' ', rightCp)`) into a
single fixed-point sum before the snap:
```cpp
return fp4::toPixel(spaceAdvanceFP + kern(leftCp, ' ') + kern(' ', rightCp));
```
This is the same single-snap pattern already used by `getTextAdvanceX`
for word widths.
### Changes
- **`GfxRenderer`**: Replaces `getSpaceKernAdjust()` with
`getSpaceAdvance()`. `getSpaceWidth()` is retained for the
single-space-word case in `measureWordWidth` where no adjacent-word kern
context is available.
- **`ParsedText`**: All four call sites (`computeLineBreaks`,
`computeHyphenatedLineBreaks`, and both loops in `extractLine`) updated
to use `getSpaceAdvance()`. The now-redundant `spaceWidth`
pre-computation and parameter are removed from all three internal layout
functions.
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**YES to analyze for
correctness**_
## Summary
Fixes a crash (load access fault) when opening EPUB chapters whose first
text block exceeds 750 words.
## Changes
* **Crash fix (`addLineToPage`)**: Added a null guard for `currentPage`.
If `makePages()` hasn't run yet (which can happen when the first block
triggers the "text block too long" split path), the page is now created
on demand.
* **Layout fix (`characterData`)**: The early-split path previously used
`viewportWidth`, ignoring CSS margins and padding. It now computes
`effectiveWidth` using `totalHorizontalInset()`, consistent with
`makePages()`.
## Additional Context
* Crash signature: `MCAUSE=0x5` (load access fault), `A0=0x0` (`Page*`
null), `MTVAL=0x4 / 0x8` (offsets into `Page::elements`).
* Confirmed in two user reports reported in #1328
* Tested on PR #1357 (not on `master`).
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**< PARTIALLY >**_
## Summary: Enable footnote anchor navigation in EPUB reader
This PR extracts the core anchor-to-page mapping mechanism from PR #1143
(TOC fragment navigation) to provide immediate footnote navigation
support. By merging this focused subset first, users get a complete
footnote experience now while simplifying the eventual review and merge
of the full #1143 PR.
---
## What this extracts from PR #1143
PR #1143 implements comprehensive TOC fragment navigation for EPUBs with
multi-chapter spine files. This PR takes only the anchor resolution
infrastructure:
- Anchor-to-page mapping in section cache: During page layout,
ChapterHtmlSlimParser records which page each HTML id attribute lands
on, serializing the map into the .bin cache file.
- Anchor resolution in `EpubReaderActivity`: When navigating to a
footnote link with a fragment (e.g., `chapter2.xhtml#note1`), the reader
resolves the anchor to a page number and jumps directly to it.
- Section file format change: Bumped to version 15, adds anchor map
offset in header.
---
## Simplified scope vs. PR #1143
To minimize conflicts and complexity, this PR differs from #1143 in key
ways:
* **Anchors tracked**
* **Origin:** Only TOC anchors (passed via `std::set`)
* **This branch:** All `id` attributes
* **Page breaks**
* **Origin**: Forces new page at TOC chapter boundaries
* **This branch:** None — natural flow
* **TOC integration**
* **Origin**: `tocBoundaries`, `getTocIndexForPage()`, chapter skip
* **This branch:** None — just footnote links
* **Bug fix**
* **This branch:** Fixed anchor page off-by-1/2 bug
The anchor recording bug (recording page number before `makePages()`
flushes previous block) was identified and fixed during this extraction.
The fix uses a deferred `pendingAnchorId` pattern that records the
anchor after page completion.
---
## Positioning for future merge
Changes are structured to minimize conflicts when #1143 eventually
merges:
- `ChapterHtmlSlimParser.cpp` `startElement()`: Both branches rewrite
the same if `(!idAttr.empty())` block. The merged version will combine
both approaches (TOC anchors get page breaks + immediate recording;
footnote anchors get deferred recording).
- `EpubReaderActivity.cpp` `render()`: The `pendingAnchor` resolution
block is positioned at the exact same insertion point where #1143 places
its `pendingTocIndex` block (line 596, right after `nextPageNumber`
assignment). During merge, both blocks will sit side-by-side.
---
## Why merge separately?
1. Immediate user value: Footnote navigation works now without waiting
for the full TOC overhaul
2. Easier review: ~100 lines vs. 500+ lines in #1143
3. Bug fix included: The page recording bug is fixed here and will carry
into #1143
4. Minimal conflicts: Structured for clean merge — both PRs touch the
same files but in complementary ways
---
### AI Usage
Did you use AI tools to help write this code? _**< YES >**_ Done by
Claude Opus 4.6
## Summary
**What is the goal of this PR?**
This change avoids the pattern of creating a `std::string` using
`.substr` in order to compare against a file extension literal.
```c++
std::string path;
if (path.length() >= 4 && path.substr(path.length() - 4) == ".ext")
```
The `checkFileExtension` utility has moved from StringUtils to
FsHelpers, to be available to code in lib/. The signature now accepts a
`std::string_view` instead of `std::string`, which makes the single
implementation reusable for Arduino `String`.
Added utility functions for commonly repeated extensions.
These changes **save about 2 KB of flash (5,999,427 to 5,997,343)**.
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**NO**_
## Summary
**What is the goal of this PR?**
Avoid building cache path strings twice, once to check existence of the
file and a second time to delete the file.
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**NO**_
## Summary
**What is the goal of this PR?** Add a user setting to decide image
support: display, show placeholder instead, supress fully
Fixes#1289
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**< NO >**_
## Summary
* **What is the goal of this PR?** Fixing two independent CSS rendering
bugs combined to make hanging-indent list styles
(e.g. margin-left:3em; text-indent:-1em) render incorrectly:
* **What changes are included?**
1. Negative text-indent was silently ignored
Three guards in ParsedText.cpp (computeLineBreaks,
computeHyphenatedLineBreaks,
extractLine) conditioned firstLineIndent on blockStyle.textIndent > 0,
so any
negative value collapsed to zero. Additionally, wordXpos was uint16_t,
which
cannot represent negative offsets — a cast of e.g. −18 would wrap to
65518 and
render the word far off-screen.
2. extraParagraphSpacing suppressed hanging indents
Even after removing the > 0 guard, the existing !extraParagraphSpacing
condition
would still suppress all text-indent when that setting is on (its
default). Positive
text-indent is a decorative paragraph indent that the user can
reasonably replace with
vertical spacing — negative text-indent is structural (it positions the
list marker)
and must always apply.
3. em unit was calibrated against line height, not font size
emSize was computed as getLineHeight() * lineCompression (the full line
advance).
CSS em units are defined relative to the font-size, which corresponds to
the
ascender height — not the line height. Using line height makes every
em-based
margin/indent ~20–30% wider than a browser would render it, and is
especially
noticeable for CSS that uses font-size: small (which we do not
implement).
## Additional Context
Test case
```
.lsl1 { margin-left: 3em; text-indent: -1em; }
<div class="lsl1">• First list item that wraps across lines</div>
<div class="lsl1">• Short item</div>
```
Before: all lines of all items started at 3 em from the left edge
(indent ignored).
After: the bullet marker hangs at 2 em; continuation lines align at 3
em.
<img width="240" alt="before"
src="https://github.com/user-attachments/assets/9dcbf3e0-fcd9-4af8-b451-a90ba4d2fb75"
/>
<img width="240" alt="after"
src="https://github.com/user-attachments/assets/1ffdcf56-a180-4267-9590-c60d7ac44707"
/>
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**YES**_
## Summary
Replaces the picojpeg library with bitbank2/JPEGDEC for JPEG decoding in
the EPUB image pipeline. JPEGDEC provides built-in coarse scaling (1/2,
1/4, 1/8), 8-bit grayscale output, and streaming block-based decoding
via callbacks.
Includes a pre-build patch script for two JPEGDEC changes affecting
progressive JPEG support and EIGHT_BIT_GRAYSCALE mode.
Closes#912
## Additional Context
# Example progressive jpeg
<img
src="https://github.com/user-attachments/assets/e63bb4f8-f862-4aa0-a01f-d1ef43a4b27a"
width="400" height="800" />
Good performance increase from JPEGDEC over picojpeg cc @bitbank2 thanks
## Baseline JPEG Decode Performance: picojpeg vs JPEGDEC (float in
callback) vs JPEGDEC (fixed-point in callback)
Tested with `test_jpeg_images.epub` on device (ESP32-C3), first decode
(no cache).
| Image | Source | Output | picojpeg | JPEGDEC float | JPEGDEC
fixed-point | vs picojpeg | vs float |
|-------|--------|--------|----------|---------------|---------------------|-------------|----------|
| jpeg_format.jpg | 350x250 | 350x250 | 313 ms | 256 ms | **104 ms** |
**3.0x** | **2.5x** |
| grayscale_test.jpg | 400x600 | 400x600 | 768 ms | 661 ms | **246 ms**
| **3.1x** | **2.7x** |
| gradient_test.jpg | 400x500 | 400x500 | 707 ms | 597 ms | **247 ms** |
**2.9x** | **2.4x** |
| centering_test.jpg | 350x400 | 350x400 | 502 ms | 412 ms | **169 ms**
| **3.0x** | **2.4x** |
| scaling_test.jpg | 1200x1500 | 464x580 | 5487 ms | 1114 ms | **668
ms** | **8.2x** | **1.7x** |
| wide_scaling_test.jpg | 1807x736 | 464x188 | 4237 ms | 642 ms | **497
ms** | **8.5x** | **1.3x** |
| cache_test_1.jpg | 400x300 | 400x300 | 422 ms | 348 ms | **141 ms** |
**3.0x** | **2.5x** |
| cache_test_2.jpg | 400x300 | 400x300 | 424 ms | 349 ms | **142 ms** |
**3.0x** | **2.5x** |
### Summary
- **1:1 scale (fixed-point vs float)**: ~2.5x faster — eliminating
software float on the FPU-less ESP32-C3 is the dominant win
- **1:1 scale (fixed-point vs picojpeg)**: ~3.0x faster overall
- **Downscaled images (vs picojpeg)**: 8-9x faster — JPEGDEC's coarse
scaling + fixed-point draw callback
- **Downscaled images (fixed-point vs float)**: 1.3-1.7x — less dramatic
since JPEG library decode dominates over the draw callback for fewer
output pixels
- The fixed-point optimization alone (vs float JPEGDEC) saved **~60% of
render time** on 1:1 images, confirming that software float emulation
was the primary bottleneck in the draw callback
- See thread for discussions on quality of progressive images,
https://github.com/crosspoint-reader/crosspoint-reader/pull/1136#issuecomment-3952952315
- and the conclusion
https://github.com/crosspoint-reader/crosspoint-reader/pull/1136#issuecomment-3959379386
- Proposal to improve quality added at
https://github.com/crosspoint-reader/crosspoint-reader/discussions/1179
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**< PARTIALLY >**_
---------
Co-authored-by: Dave Allie <dave@daveallie.com>
## Summary
**What is the goal of this PR?** Replace the linear scan of
`lookupHtmlEntity` with a simple binary search to improve lookup
performance.
**What changes are included?**
`lib/Epub/Epub/Entities/htmlEntities.cpp`:
- Sorted the `ENTITY_LOOKUP` array.
- Added a compile-time assertion to guarantee the array remains sorted.
- Rewrote `lookupHtmlEntity` to use a binary search.
## Additional Context
Benchmarked on my x64 laptop (probably will be different on RISC-V)
```
=== Benchmark (53 entities x 10000 iterations) ===
Version Total time Avg per lookup
----------------------------------------------
linear 236.97 ms total 447.11 ns/lookup
binary search 22.09 ms total 41.68 ns/lookup
=== Summary ===
Binary search is 10.73x faster than linear scan.
```
This is a simplified alternative to #1180, focused on keeping the
implementation clean, and maintainable.
### AI Usage
Did you use AI tools to help write this code? _**< NO >**_
---------
Co-authored-by: Zach Nelson <zach@zdnelson.com>
## Summary
**What is the goal of this PR?** Implement support for footnotes in epub
files.
It is based on #553, but simplified — removed the parts which
complicated the code and burden the CPU/RAM. This version supports basic
footnotes and lets the user jump from location to location inside the
epub.
**What changes are included?**
- `FootnoteEntry` struct — A small POD struct (number[24], href[64])
shared between parser, page storage, and UI.
- Parser: `<a href>` detection (`ChapterHtmlSlimParser`) — During a
single parsing pass, internal epub links are detected and collected as
footnotes. The link text is underlined to hint navigability.
Bracket/whitespace normalization is applied to the display label (e.g.
[1] → 1).
- Footnote-to-page assignment (`ChapterHtmlSlimParser`, `Page`) —
Footnotes are attached to the exact page where their anchor word
appears, tracked via a cumulative word counter during layout, surviving
paragraph splits and the 750-word mid-paragraph safety flush.
- Page serialization (`Page`, `Section`) — Footnotes are
serialized/deserialized per page (max 16 per page). Section cache
version bumped to 14 to force a clean rebuild.
- Href → spine resolution (`Epub`) — `resolveHrefToSpineIndex()` maps an
href (e.g. `chapter2.xhtml#note1`) to its spine index by filename
matching.
- Footnotes menu + activity (`EpubReaderMenuActivity`,
`EpubReaderFootnotesActivity`) — A new "Footnotes" entry in the reader
menu lists all footnote links found on the current page. The user
scrolls and selects to navigate.
- Navigate & restore (`EpubReaderActivity`) — `navigateToHref()` saves
the current spine index and page number, then jumps to the target. The
Back button restores the saved position when the user is done reading
the footnote.
**Additional Context**
**What was removed vs #553:** virtual spine items
(`addVirtualSpineItem`, `isVirtualSpineItem`), two-pass parsing,
`<aside>` content extraction to temp HTML files, `<p class="note">`
paragraph note extraction, `replaceHtmlEntities` (master already has
`lookupHtmlEntity`), `footnotePages` / `buildFilteredChapterList`,
`noterefCallback` / `Noteref` struct, and the stack size increase from 8
KB to 24 KB (not needed without two-pass parsing and virtual file I/O on
the render task).
**Performance:** Single-pass parsing. No new heap allocations in the hot
path — footnote text is collected into fixed stack buffers (char[24],
char[64]). Active runtime memory is ~2.8 KB worst-case (one page × 16
footnotes × 88 bytes, mirrored in `currentPageFootnotes`). Flash usage
is unchanged at 97.4%; RAM stays at 31%.
**Known limitations:** When clicking a footnote, it jumps to the start
of the HTML file instead of the specific anchor. This could be
problematic for books that don't have separate files for each footnote.
(no element-id-to-page mapping yet - will be another PR soon).
---
### AI Usage
Did you use AI tools to help write this code? _**< PARTIALLY>**_
Claude Opus 4.6 was used to do most of the migration, I checked manually
its work, and fixed some stuff, but I haven't review all the changes
yet, so feedback is welcomed.
---------
Co-authored-by: Arthur Tazhitdinov <lisnake@gmail.com>
## Summary
* **What is the goal of this PR?** KOReader sync on a German-language
book would fail with an out-of-memory error when trying to open the
destination chapter after applying remote progress. The root cause was a
chain of two independent bugs that combined to exhaust the contiguous
heap needed by the EPUB inflate pipeline.
* **What changes are included?**
## Fix 1 — Hyphenation heap defragmentation (LiangHyphenation.cpp)
### What was happening
AugmentedWord, the internal struct used during Liang pattern matching,
held three std::vector<> members (bytes, charByteOffsets,
byteToCharIndex) plus a separate scores vector — a total of 4 heap
allocations per word during page layout. For a German-language section
with hundreds of words, thousands of small malloc/free cycles fragmented
the heap. Total free memory was adequate (~108 KB) but the largest
contiguous block shrank well below the 32 KB needed for the INFLATE ring
buffer used during EPUB decompression. The failure was invisible with
hyphenation disabled, where MaxAlloc stayed at ~77 KB; enabling German
hyphenation silently destroyed the contiguity the allocator needed.
### What changed
The three std::vector<> members of AugmentedWord and the scores vector
are replaced with fixed-size C arrays on the render-task stack:
```
uint8_t bytes[160] // was std::vector<uint8_t>
size_t charByteOffsets[70] // was std::vector<size_t>
int32_t byteToCharIndex[160] // was std::vector<int32_t>
uint8_t scores[70] // was std::vector<uint8_t> (local in liangBreakIndexes)
```
Sizing is based on the longest known German word (~63 codepoints × 2
UTF-8 bytes + 2 sentinel dots = 128 bytes); MAX_WORD_BYTES=160 and
MAX_WORD_CHARS=70 give comfortable headroom. The same analysis holds for
all seven supported languages (en, fr, de, es, it, ru, uk) — every
accepted letter encodes to at most 2 UTF-8 bytes after case-folding.
Words exceeding the limits are silently skipped (no hyphenation
applied), which is correct behaviour. The struct lives on the 8 KB
render-task stack so no permanent DRAM is consumed.
Verification: after the fix, MaxAlloc reads 77,812 bytes with German
hyphenation enabled — identical to the figure previously achievable only
with hyphenation off.
## Fix 2 — WiFi lifecycle in KOReaderSyncActivity
(KOReaderSyncActivity.cpp)
### What was happening
onEnter() called WiFi.mode(WIFI_STA) unconditionally before delegating
to WifiSelectionActivity. WifiSelectionActivity manages WiFi mode
internally (it calls WiFi.mode(WIFI_STA) again at scan start and at
connection attempt). The pre-emptive call from KOReaderSyncActivity
interfered with the sub-activity's own state machine, causing
intermittent connection failures that were difficult to reproduce.
Additionally, WiFi was only shut down in onExit(). If the user chose
"Apply remote progress" the activity exited without turning WiFi off
first, leaving the radio on and its memory allocated while the EPUB was
being decompressed — unnecessarily consuming the contiguous heap
headroom that inflate needed.
### What changed
* WiFi.mode(WIFI_STA) removed from onEnter(). WifiSelectionActivity owns
WiFi mode; KOReaderSyncActivity should not touch it before the
sub-activity runs.
* A wifiOff() helper (SNTP stop + disconnect + WIFI_OFF with settling
delays) is extracted into the anonymous namespace and called at every
web-session exit point:
- "Apply remote" path in loop() — before onSyncComplete()
- performUpload() success path
- performUpload() failure path
- onExit() (safety net for all other exit paths)
## Additional Context
* Add any other information that might be helpful for the reviewer
(e.g., performance implications, potential risks,
specific areas to focus on).
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**YES**_ and two days of
blood, sweat and heavy swearing...
## Summary
**What is the goal of this PR?**
Improved typesetting, including
[kerning](https://en.wikipedia.org/wiki/Kerning) and
[ligatures](https://en.wikipedia.org/wiki/Ligature_(writing)#Latin_alphabet).
**What changes are included?**
- The script to convert built-in fonts now adds kerning and ligature
information to the generated font headers.
- Epub page layout calculates proper kerning spaces and makes ligature
substitutions according to the selected font.
## Additional Context
- I am not a typography expert.
- The implementation has been reworked from the earlier version, so it
is no longer necessary to omit Open Dyslexic, and kerning data now
covers all fonts, styles, and codepoints for which we include bitmap
data.
- Claude Opus 4.6 helped with a lot of this.
- There's an included test epub document with lots of kerning and
ligature examples, shown in the photos.
**_After some time to mature, I think this change is in decent shape to
merge and get people testing._**
After opening this PR I came across #660, which overlaps in adding
ligature support.
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**YES, Claude Opus 4.6**_
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
## Summary
Fix redefinition of `FILE_*` macro.
Note that there will still be 2 warning:
```
.pio/libdeps/default/WebSockets/src/WebSocketsClient.cpp: In member function 'void WebSocketsClient::clientDisconnect(WSclient_t*, const char*)':
.pio/libdeps/default/WebSockets/src/WebSocketsClient.cpp:573:31: warning: 'virtual void NetworkClient::flush()' is deprecated: Use clear() instead. [-Wdeprecated-declarations]
573 | client->tcp->flush();
| ~~~~~~~~~~~~~~~~~~^~
```
--> I assume the upstream library need to fix it
And:
```
src/activities/Activity.cpp:8:1: warning: 'noreturn' function does return
8 | }
| ^
```
Will be fixed in #1016
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? **NO**
## Summary
* **What is the goal of this PR?** Address expected hyphenation issue
from
https://github.com/crosspoint-reader/crosspoint-reader/issues/998#issuecomment-3940533510
* Closes#998
* **What changes are included?** Add `„` (U+201E, _Double Low-9
Quotation Mark_), `‚` (U+201A, _Single Low-9 Quotation Mark_) and `‹`
(U+2039, _Single Left-pointing Angle Quotation Mark_) exceptions, other
quote types were handled correctly.
**Before**
<img width="480" height="155" alt="hyph3"
src="https://github.com/user-attachments/assets/e06b4071-2c8c-4814-965d-96fbe302a450"
/>
**After**
<img width="480" height="154" alt="hyph3-fix"
src="https://github.com/user-attachments/assets/4f7f5406-e200-451c-8bee-3f410cc84bbe"
/>
## Additional Context
* Add any other information that might be helpful for the reviewer
(e.g., performance implications, potential risks,
specific areas to focus on).
---
### AI Usage
While CrossPoint doesn't have restrictions on AI tools in contributing,
please be transparent about their usage as it
helps set the right context for reviewers.
Did you use AI tools to help write this code? _**NO**_
## Summary
_Revision to @blindbat's #802. Description comes from the original PR._
- Replace `std::list` with `std::vector` for word storage in `TextBlock`
and `ParsedText`
- Use index-based access (`words[i]`) instead of iterator advancement
(`std::advance(it, n)`)
- Remove the separate `continuesVec` copy that was built from
`wordContinues` for O(1) access — now unnecessary since
`std::vector<bool>` already provides O(1) indexing
## Why
`std::list` allocates each node individually on the heap with 16 bytes
of prev/next pointer overhead per node. For text layout with many small
words, this means:
- Scattered heap allocations instead of contiguous memory
- Poor cache locality during iteration (each node can be anywhere in
memory)
- Per-node malloc/free overhead during construction and destruction
`std::vector` stores elements contiguously, giving better cache
performance during the tight rendering and layout loops. The
`extractLine` function also benefits: list splice was O(1) but required
maintaining three parallel iterators, while vector range construction
with move iterators is simpler and still efficient for the small
line-sized chunks involved.
## Files changed
- `lib/Epub/Epub/blocks/TextBlock.h` / `.cpp`
- `lib/Epub/Epub/ParsedText.h` / `.cpp`
## AI Usage
YES
## Test plan
- [ ] Open an EPUB with mixed formatting (bold, italic, underline) —
verify text renders correctly
- [ ] Open a book with justified text — verify word spacing is correct
- [ ] Open a book with hyphenation enabled — verify words break
correctly at hyphens
- [ ] Navigate through pages rapidly — verify no rendering glitches or
crashes
- [ ] Open a book with long paragraphs — verify text layout matches
pre-change behavior
---------
Co-authored-by: Kuanysh Bekkulov <kbekkulov@gmail.com>
Dave Allie
Justin MitchellandGitHub
martin brookandGitHub
jpirnayandGitHub
Zach NelsonandGitHub
Adrian Wilkins-CaruanaandGitHub
Uri TauberandGitHub
Xuan-Son NguyenandGitHub
Florian HülsmannandGitHub
