Move third-party sources, tighten encapsulation, extend Konsist rules

[lemcoder]

Two commits.

1. Move vendored checkouts under third-party/

markitdown and anydoc are reference sources for the benchmark, not part of the build. third-party/markitdown and third-party/anydoc now; the repository root is this project's own layout again. scripts/benchmark.py also picks up a cargo build --release inside third-party/anydoc.

2. Encapsulation, and rules that keep it

The duplication that kept biting. JVM and Android held nine byte-identical converter files. They drifted twice during recent work — once in EpubConverter, once in PptxConverter — and nothing caught it either time. A jvmShared source set now holds every converter both targets run. Only PdfConverter (pdfbox vs pdfbox-android) and the platform factories remain per-target.

library/src/
  commonMain/    model, render, pipeline, plain-text converters
  jvmShared/     csv json xml html epub docx xlsx pptx + HtmlToDocument   ← was duplicated
  jvmMain/       PdfConverter, TikaMimeDetector, factory
  androidMain/   PdfConverter, AndroidMimeDetector, factory

Explicit API mode. The library now builds with explicitApi(), so every exported declaration states its visibility and return type. This surfaced several helpers that were public only because nobody wrote a modifier: HtmlToDocument, plainTextBlocks, TikaMimeDetector and AndroidMimeDetector are all internal now.

Smaller cleanups.

• converter selection moved into a ConverterRegistry that sorts on registration and never exposes its entries, replacing a Pair<DocumentConverter, Double> list re-sorted on every parse
• DocumentBuilder.title is private behind titleIfAbsent() / title()
• both builders carry a @DocumentDsl marker, so an inlines { } block cannot silently call document-level methods

Six new Konsist rules (11 total), each verified to fail against a planted violation before being committed:

rule	catches
helpers under utils are never public	implementation detail leaking into the API
the model exposes no mutable state	var on an exported model type
every DocumentConverter is named *Converter and lives in converters	converters growing outside their package
converters do not import each other	conversion chaining behind the pipeline's back
no wildcard imports, no printing from library code	hygiene
no production file in two source sets	the duplication above, coming back

The last one names the two legitimately platform-specific files explicitly, so adding a third requires saying so out loud.

Verification

./gradlew check passes: JVM tests, Android host tests, CLI build, detekt, ktfmtCheck, and all 11 architecture rules.

Converted output is byte-identical across all eight fixtures — this refactor moves code without changing behaviour.

🤖 Generated with Claude Code