<!-- CURSOR_SUMMARY -->
> [!NOTE]
> **Medium Risk**
> Introduces a new automated publishing workflow and modifies
dependency-install semantics in CI/Docker, which could cause release or
build failures if credentials, tags, or lockfile expectations are
misconfigured.
> 
> **Overview**
> Adds an automated release pipeline: a new `release.yml` workflow
triggers on published GitHub releases, validates the tag matches
`unstructured.__version__`, builds via `uv build`, publishes to PyPI
using trusted publishing, and *best-effort* uploads the same artifacts
to Azure Artifacts via `twine`.
> 
> Across CI, Docker, and Make targets, replaces `uv sync --frozen` with
`uv sync --locked` and adds `uv run --no-sync` where `uv sync` already
ran to avoid implicit re-syncing; introduces a new `release` dependency
group (adds `twine`), bumps version to `0.20.2`, and updates `uv.lock`
accordingly.
> 
> <sup>Written by [Cursor
Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit
c9555c9. This will update automatically
on new commits. Configure
[here](https://cursor.com/dashboard?tab=bugbot).</sup>
<!-- /CURSOR_SUMMARY -->

#4215)

Closes #3864

## Summary
- Fixes issue where bold text in PDFs is extracted with duplicate
characters (e.g., "BOLD" → "BBOOLLDD")
- Some PDF generators simulate bold by rendering each character twice at
slightly offset positions
- Added character-level deduplication based on position proximity to
detect and remove these duplicates

## Problem
When extracting text from certain PDFs, bold text appears duplicated:
```python
# Before fix
elements = partition_pdf("document.pdf", strategy="fast")
print(elements[0].text)  # Output: ">60>60" instead of ">60"
```

## Solution
Added character-level deduplication that:
- Compares consecutive characters' text content and position
- Removes duplicates where same character appears within 3 pixels
(configurable)
- Preserves spaces and other non-character elements (LTAnno objects)

```python
# After fix
elements = partition_pdf("document.pdf", strategy="fast")
print(elements[0].text)  # Output: ">60" ✓
```

## Configuration
```bash
# Default: 3.0 pixels (enabled)
export PDF_CHAR_DUPLICATE_THRESHOLD=3.0

# Disable deduplication
export PDF_CHAR_DUPLICATE_THRESHOLD=0

# More aggressive deduplication
export PDF_CHAR_DUPLICATE_THRESHOLD=5.0
```

Improve PDF fast strategy cold-start latency by lazy-loading hi-res-only
imports in
[pdf.py](https://github.com/Unstructured-IO/unstructured/blob/1c3d5e6ef7b6123a2d8739bf9a8c3afecc3dd127/unstructured/partition/pdf.py).

This reduces first-call startup overhead without changing partition
behavior.

Local benchmarks show a significant fast strategy cold-start speedup of
~35% from 2.75s -> 1.78s.
They also show a small hi_res slowdown (~2-4%), which is acceptable
given the fast improvements.

Benchmark was run on 6 pdfs

https://github.com/Unstructured-IO/unstructured/blob/main/example-docs/pdf/DA-1p.pdf

https://github.com/Unstructured-IO/unstructured/blob/main/example-docs/pdf/chevron-page.pdf

https://github.com/Unstructured-IO/unstructured/blob/main/example-docs/pdf/embedded-images-tables.pdf

https://github.com/Unstructured-IO/unstructured/blob/main/example-docs/pdf/fake-memo-with-duplicate-page.pdf

https://github.com/Unstructured-IO/unstructured/blob/main/example-docs/pdf/interface-config-guide-p93.pdf

https://github.com/Unstructured-IO/unstructured/blob/main/example-docs/pdf/layout-parser-paper-fast.pdf

<!-- CURSOR_SUMMARY -->
---

> [!NOTE]
> **Medium Risk**
> Touches core PDF partitioning by changing import timing and locations;
behavior should be unchanged but there is some risk of
missed/conditional imports causing runtime errors in less-tested
hi_res/OCR/analysis paths.
> 
> **Overview**
> Improves PDF `fast` strategy cold-start performance by **lazy-loading
hi-res-only dependencies** in `unstructured/partition/pdf.py` (moving
several `pdf_image`/`unstructured_inference`-related imports into
`_partition_pdf_or_image_local` and other hi-res/OCR-only code paths),
while keeping the `fast` path lighter.
> 
> Adds `scripts/performance/quick_partition_bench.py` for quick local
cold vs warm partition timing across one or more PDFs, updates the table
metrics helper to import `convert_pdf_to_images` from `pdf_image_utils`,
and bumps the library version to `0.20.4` with corresponding changelog
entry.
> 
> <sup>Written by [Cursor
Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit
b66ae0e. This will update automatically
on new commits. Configure
[here](https://cursor.com/dashboard?tab=bugbot).</sup>
<!-- /CURSOR_SUMMARY -->

This PR fixes an issue where an invalid `text_as_html` input into html
based table chunking logic can lead to chunking failing. Like the
following stack trace shows:

```
    |   File "/app/unstructured/unstructured/chunking/base.py", line 594, in iter_chunks
    |     yield from _TableChunker.iter_chunks(
    |   File "/app/unstructured/unstructured/chunking/base.py", line 837, in _iter_chunks
    |     html_size = measure(self._html) if self._html else 0
    |                                        ^^^^^^^^^^
    |   File "/app/unstructured/unstructured/utils.py", line 154, in __get__
    |     value = self._fget(obj)
    |             ^^^^^^^^^^^^^^^
    |   File "/app/unstructured/unstructured/chunking/base.py", line 866, in _html
    |     if not (html_table := self._html_table):
    |                           ^^^^^^^^^^^^^^^^
    |   File "/app/unstructured/unstructured/utils.py", line 154, in __get__
    |     value = self._fget(obj)
    |             ^^^^^^^^^^^^^^^
    |   File "/app/unstructured/unstructured/chunking/base.py", line 884, in _html_table
    |     return HtmlTable.from_html_text(text_as_html)
    |            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    |   File "/app/unstructured/unstructured/common/html_table.py", line 61, in from_html_text
    |     root = fragment_fromstring(html_text)
    |            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    |   File "/app/.venv/lib/python3.12/site-packages/lxml/html/__init__.py", line 810, in fragment_fromstring
    |     elements = fragments_fromstring(
    |                ^^^^^^^^^^^^^^^^^^^^^
    |   File "/app/.venv/lib/python3.12/site-packages/lxml/html/__init__.py", line 780, in fragments_fromstring
    |     raise etree.ParserError(
    | lxml.etree.ParserError: There is leading text: '```html\n'
```

The solution is to catch the parser error and return a `None` instead in
`unstructured/chunking/base.py` in `_html_table`. This way we fallback
to text based chunking for this element with a warning log.

This PR addresses an issue where hashing element id loses the reference
for parent id.
This happens when calling `partition_html` where the partition process
already assigned parent ids for elements based on html structure before
`apply_metadata` is called, i.e., before element id hashing happens.
This fix ensures that the parent references stay unchanged after
hashing.