docs: refresh CLAUDE.md for grammar/structured-output, perplexity, KVarN defaults

[pekkah]

Summary

Updates the checked-in CLAUDE.md to reflect features added since it was last written. Documentation-only change — no code touched.

Changes

• Grammar / structured-output subsystem (feat(core): caller-supplied output-level ITokenConstraint (whole-turn grammar, not just tool args) #423/JsonSchemaOutputConstraint: enforce declared property order (blocks streaming an early field before a later one) #425) — documents the new Core/Grammar/ directory: ITokenConstraint, JsonSchemaOutputConstraint (whole-turn JSON-schema output), the per-family tool-argument constraints, and ToolSchemaCompiler. Adds ITokenConstraint to the central interfaces list and notes the Sampler's per-step constraint masking.
• perplexity CLI command — adds PerplexityCommand (KVarN accuracy gate, Implement KVarN variance-normalized KV-cache quantization (2-bit values) #180) to the command examples and the CLI frontend description.
• JSON-schema structured output flags — -j/--json-schema/--json-schema-file/--json-schema-ordered on RunCommand, with a usage example and the note that the server exposes the same via response_format:json_schema.
• KVarN / --tq-mode auto (fix(tq): default --tq-mode to auto — KVarN where supported, Lloyd-Max fallback with quality warning (#432) #436/TurboQuant Lloyd-Max 3-bit collapses on Qwen3 (PPL 61x, degenerate decode) — pre-existing on master #432/Implement KVarN variance-normalized KV-cache quantization (2-bit values) #180) — clarifies the auto default (KVarN where supported, Lloyd-Max fallback with a quality warning) and the CPU (AVX2) + CUDA-graph decode paths.
• SnapKvEval benchmark project — added to Samples & Scripts; grammar coverage noted in the Tests.Core table row.

Test plan

Documentation-only; no build or test impact.

🤖 Generated with Claude Code

---

Generated by Claude Code

[gemini-code-assist]

Code Review

This pull request updates the CLAUDE.md documentation to reflect several new features, including perplexity evaluation over a corpus, whole-turn structured JSON-schema output, grammar-constrained decoding, and the new auto mode for TurboQuant KV cache compression. The review feedback suggests correcting the spelling of 'honours' to 'honors' in the CLI documentation to maintain consistency with the American English spelling used throughout the rest of the codebase.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

To maintain spelling consistency across the codebase, please use the American English spelling honors instead of the British English spelling honours. The rest of the codebase consistently uses American English (e.g., honored in Sampler.cs and RunCommand.cs, and honor in RunCommand.cs).

Suggested change:

   - **CLI** (`SharpInference.Cli`, `Spectre.Console.Cli`, llama.cpp-compatible flags): `RunCommand` (default text/vision inference; also whole-turn JSON-schema structured output via `-j`/`--json-schema`/`--json-schema-file`), `ImageCommand` (`image` subcommand), `PerplexityCommand` (`perplexity` — accuracy gate over a corpus, honors `--tq`/`--tq-mode`), `ListMetadataCommand` (`list-metadata`), `ListTensorsCommand` (`list-tensors`).