AutoChunking + Performance Enhancements

[techbuzzz]

Pull Request

Description

v10.2.3 — AutoChunking + Performance Enhancements

This PR delivers two major bodies of work:

1. ✂️ AutoChunking — Progressive Disclosure for Large Documents

When a knowledge document (contract, report, article) exceeds the smallest model's context window, Delibera now automatically splits it into chunks and distributes them across debate rounds via progressive disclosure. Every model receives a complete view of the document by the final round — no more silent context truncation or API errors.

Key additions:

• Chunking namespace — AutoChunker (3 strategies: SemanticBoundary, FixedSize, SlidingWindow), AutoChunkingOrchestrator, AutoChunkingOptions
• ModelCapabilities record + ModelContextWindowRegistry (40+ pre-registered models: Llama, Qwen, DeepSeek, Phi, Mistral, Gemma, GPT, Claude, YandexGPT…)
• ILLMProvider.GetModelCapabilitiesAsync() — Ollama queries /api/show for num_ctx; other providers fall back to the registry
• PromptContext extended with ChunkingPlan, AutoChunkingEnabled, MinContextWindow, GetChunkedUserPrompt()
• All debate strategies (StandardDebate, CritiqueDebate, ConsensusDebate) use BuildChunkedPrompt() for round-aware chunk distribution
• CouncilBuilder.WithOptions() — three configuration paths: fluent API, CouncilOptions snapshot, inline lambda. CouncilBuilder(CouncilOptions) constructor for one-shot setup
• DI auto-wiring — IOptions<CouncilOptions> → CouncilBuilder constructor automatically
• AutoChunkingConfig in CouncilOptions with full appsettings.json binding
• Console example — dotnet run -- --autochunking

2. ⚡ Performance Enhancements — High Availability & High Performance

Full audit of the hot paths across the entire codebase. 9 fixes applied:

#	Area	Problem	Fix	Impact
1	AutoChunker.SplitByHeaders	O(n·k) per iteration (5× IndexOf in while-loop)	Single-pass ReadOnlySpan<char> scan	O(n)
2	SemanticCompressor.SplitSentences	O(n·k) per iteration (6× IndexOf in while-loop)	Single-pass char scan	O(n)
3	AutoChunkingOrchestrator	Sequential model capability queries	Task.WhenAll — parallel queries	N→1 round-trip
4	AutoChunker.SplitBySeparators	string.Split() array allocations	Span-based IndexOf + manual split	0 allocations
5	KnowledgeKeeper chunk counting	context.Split("[Source ").Length - 1	CountOccurrences() — IndexOf loop	0 allocations
6	PromptContext.GetChunkedUserPrompt	.Skip().Take().ToList() LINQ allocations	Index-based for loop	0 enumerator allocations
7	HybridCompressor	new DeduplicationCompressor() etc. on every call	Lazy-cached _dedup/_semantic/_summarizer fields	0 allocations on reuse
8	TokenCounter memoization	No eviction — new values silently dropped when full	LRU eviction with LinkedList<string> + batch eviction (64)	Correct cache under load
9	CouncilBuilder.ApplyOptions	Magic strings/numbers	Named constants DefaultStrategy, DefaultMaxRounds, DefaultTemperature	Readability + safety

Already-optimized code confirmed correct:

• CompressionCache — ArrayPool<T>, stackalloc, ReaderWriterLockSlim, batch eviction
• SemanticCompressor.CosineSimilarity — SIMD Vector<float> hardware acceleration
• DeduplicationCompressor — CollectionsMarshal.AsSpan(), batch processing (16)
• TokenCounter.EstimateTokens(ReadOnlySpan<char>) — allocation-free
• OllamaProvider — Polly.ResilienceContextPool.Shared

Documentation

• CHANGELOG.md — full [10.2.3] section
• README.md / README-RU.md — new AutoChunking section (how it works, 3 config paths, appsettings, model registry)
• src/Delibera.Core/README.md (NuGet) — AutoChunking section + updated appsettings example
• Delibera.Core.csproj — version 10.2.2 → 10.2.3, updated Description, PackageTags, PackageReleaseNotes

Related Issues

N/A — feature-driven release.

Type of Change

• ✨ New feature (non-breaking change that adds functionality)
• ⚡ Performance enhancement
• 📚 Documentation update

Checklist

• My code follows the project's coding standards (see CONTRIBUTING.md)
• The solution builds with 0 errors and 0 warnings (dotnet build --configuration Release)
• I have added/updated XML documentation for public APIs
• I have updated relevant documentation (README.md, README-RU.md, CHANGELOG.md, NuGet README)
• I have tested my changes (examples run as expected)

Additional Notes

• Fully backward compatible — AutoChunking is opt-in (disabled by default). All existing ILLMProvider implementations continue to work (default GetModelCapabilitiesAsync returns null). PromptContext is a record with with-expression support.
• 31 files changed, +2,899 / −319 lines.
• The --autochunking console example includes an offline mode that demonstrates chunking plans for 4K/8K/32K/128K context windows without requiring a running Ollama instance.