Feat/v0.2.0 model foundation

[snss10]

Summary

• Introduces vendor-neutral model-family adapters for prompts, sampling, stop sequences, and output cleanup.
• Preserves the existing Qwen 2.5 model and first-launch download behavior.
• Adds a development-only qualification corpus and runner for comparing local GGUF models.
• Adds Windows hardware inspection for memory, storage, architecture, processor count, and AVX2 support.
• Adds hardware-based model recommendation using measured quality, memory, storage, backend, and responsiveness requirements.
• Adds atomic persistence for Automatic/Manual model selection and the active model ID.
• Recovers safely from missing, corrupt, outdated, or unknown model settings.

Scope

• No additional model is exposed to users yet.
• No model is bundled with the application.
• The existing Qwen model remains the application default.
• Model-management UI and model switching will be added in later phases.
• Benchmark results and local GGUF files are not committed or published.

Validation

• dotnet restore TextRecast.slnx
• dotnet format TextRecast.slnx --no-restore --verify-no-changes
• dotnet test TextRecast.slnx -c Release --no-restore
• dotnet build TextRecast.slnx -c Release --no-restore
• Self-contained win-x64 publish completed successfully.
• 52 automated tests passed.
• Release build completed with 0 warnings and 0 errors.
• Published executable startup smoke test passed.
• Published package contains the executable, runtime, and required legal files.
• Published package contains no GGUF models or benchmark tooling.

Current model baseline

A local qualification run against Qwen 2.5 1.5B completed all 18 cases:

• Overall score: 9.0/10
• No inference failures
• No protocol leakage
• No repetition failures
• English and Japanese cases performed well
• Hindi, Spanish, French, and German inputs were translated into English

The multilingual result is an existing model limitation, not a regression introduced by this PR. The new benchmark tooling correctly identifies it and will help select better models during the qualification phase.

Foreground end-to-end validation

A complete foreground workflow was tested using a controlled WPF editor:

• Selected text was captured successfully.
• The real local model generated the expected correction.
• The selected text was replaced successfully.
• Clipboard contents were restored after replacement.
• TextRecast remained responsive throughout the operation.
• End-to-end formatting and replacement completed in 4.79 seconds.

Input:

teh report dont include the final deadline

Replacement:

The report does not include the final deadline.