## Summary
**Goal:** Fix bidirectional KOSync position matching between CrossPoint
and KOReader so that syncing in either direction lands on the correct
page with character-level accuracy.
**Changes included:**
**Download — `toCrossPoint` (server XPath → CrossPoint page)**
- **XPath ancestry mode for structured elements**: The previous
`ParagraphStreamer` only tracked `<p>` elements. Replaced with a full
ancestor-walking mode that correctly resolves XPaths pointing into
`<li>`, `<ul>`, and other structured elements. Char offset within the
target element is bounded to the matched element's content only.
- **Slash-in-attribute-value corrupts depth tracking**:
`processByteInTag()` treated every `/` byte as a self-closing tag
marker, including `/` inside quoted attribute values (e.g.
`xmlns="http://..."`, `src="Links/image.jpg"`). This drove `htmlDepth`
to 0 prematurely, causing the ancestry search to exit far short of the
target paragraph. Fixed with `inAttrQuote` tracking.
- **Off-by-one in page formula**: `intra * totalPages` rounds up
incorrectly for last-page positions. Changed to `intra * (totalPages -
1)` to map the `[0, 1]` intra fraction correctly onto the `[0,
totalPages-1]` page range. Example: page 14 of 17 was returned as 15.
**Upload — `toKOReader` (CrossPoint page → server XPath)**
- **Off-by-one in page-to-intra formula**: Symmetric fix — `pageNumber /
totalPages` changed to `pageNumber / (totalPages - 1)`, with the guard
updated from `> 0` to `> 1` to avoid division by zero.
- **`<li>`-based XPath generation**: When the current page starts on a
list item, `findXPathForProgress` now generates `ul[N]/li[M]` XPaths
rather than falling back to the preceding `<p>`. Requires the new
`listItemIndex` field in `PageLutEntry` (section cache version bumped to
23).
- **Text-node precision with correct `text()[N].M` format**: KOReader
expects `text()[N].M` where `N` is the 1-based index of the specific
text node within the element. The previous attempt generated `text().M`
(no brackets), which caused KOReader to jump to the front of the book.
Implements a per-element text-node index stack in
`XPathProgressResolver` — parallel to the existing element path stack —
that correctly tracks text node indices relative to each element. Empty
text nodes from bare anchor elements (`<a id="anchor"/>`) are
intentionally skipped, matching KOReader's own text node counting
behavior.
**Reviewer-caught bugs**
- **Double `onCloseTag()` on malformed `</br/>`**: Both the `tagIsClose`
path and the self-closing `/` check were firing, double-decrementing
`htmlDepth`. Fixed with a `!tagIsClose` guard.
- **Dangling pointer in `LOG_DBG`**:
`std::to_string(*nextParagraphPage).c_str()` passed a pointer to a
temporary destroyed before the variadic call. Fixed with `snprintf` into
a stack `char[8]` buffer.
## Additional Context
- Section cache version bumped from 22 → 23 due to the new
`listItemIndex` field in `PageLutEntry`. Users upgrading will see a
one-time re-render of all cached sections on first load — no data loss.
- The `textNodeIndexStack` in `XPathProgressResolver` is a
`std::vector<int>` that mirrors the existing `path` and `parentStates`
stacks — same depth, same lifetime. No additional heap pressure beyond
what was already present.
- All fixes verified on device with *Gentle and Lowly* by Dane C.
Ortlund (spine 21, 17 pages). Download syncs land on the correct page;
upload syncs land at the correct paragraph with character-level offset.
## Test plan
- [ ] Download: sync from KOReader → CrossPoint lands on correct page
for `text()[N].M` XPaths
- [ ] Download: ancestry correctly resolves `<li>` positions inbound
from KOReader
- [ ] Upload: sync from CrossPoint → KOReader lands within one page for
mid-paragraph positions
- [ ] Upload: sync from CrossPoint → KOReader correctly targets `<li>`
elements when page starts on a list item
- [ ] Upload: `text()[N].M` format XPaths do not cause KOReader to jump
to front of book
- [ ] Section cache version 23: delete `.crosspoint/` and verify clean
re-parse with no crashes
---
### AI Usage
Did you use AI tools to help write this code? **YES** — developed with
Claude Code (Anthropic).
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
