Expand CUDA/PTX docs, improve retrieval aliases, and add search regression suite

[qhy991]

Summary

Add comprehensive CUDA/PTX documentation and retrieval-quality infrastructure for Context Hub, with English-only content, richer PTX reference guidance, search alias tags, and a runnable search regression suite.

Changes

Documentation Expansion

• Added many CUDA topic docs under content/cuda/docs/* (shared memory, synchronization, warp primitives, WMMA/Tensor Core, CUDA Core, performance and production playbooks).
• Added PTX pattern docs:
  • content/cuda/docs/ptx-integer-bit-manipulation-patterns/DOC.md
  • content/cuda/docs/ptx-atomic-and-reduction-patterns/DOC.md
  • content/cuda/docs/ptx-mbarrier-protocol-patterns/DOC.md
  • content/cuda/docs/ptx-warp-synchronization-patterns/DOC.md
• Expanded a large set of PTX reference pages with practical sections such as Usage Notes, Common Failure Modes, and protocol checklists.

Retrieval Improvements

• Standardized PTX top-level doc language metadata to support --lang cpp retrieval paths while preserving PTX semantics.
• Added search-friendly alias/synonym tags to high-frequency docs (WMMA/Tensor Core/shared memory/coalescing/synchronization/atomics/mbarrier topics), including terms like:
  • warp-matrix-multiply-accumulate, tensorcore, smem, coalesced-access, cas-loop, syncthreads.

Search Regression Suite

• Added runnable regression tooling:
  • scripts/search_regression.py
  • scripts/search_regression_cases.json
  • scripts/search_regression_baseline.json
  • docs/features/search-regression.md
• Includes 18 regression queries with top-k expectations for WMMA/shared-memory/tensor-core/PTX and optimization workflows.

What

This PR upgrades Context Hub from “basic CUDA/PTX coverage” to a retrieval-oriented, regression-tested documentation set designed for agent use.

Why

Problems addressed:

• Sparse or uneven PTX reference pages reduced practical usefulness for agent-generated kernels.
• Query phrasing variance (aliases/synonyms) could reduce retrieval quality.
• No stable search regression gate existed to detect ranking drift after content updates.
• Prior mixed-language artifacts and metadata mismatches could hurt consistency in local retrieval flows.

Testing

Content/Build Validation

• ./cli/bin/chub build content --validate-only -> PASSED (1622 docs, 6 skills, 0 warnings)
• ./cli/bin/chub build content -> PASSED

Search Regression

• python3 scripts/search_regression.py --mode check -> PASSED (18/18)
• python3 scripts/search_regression.py --mode snapshot --snapshot-out scripts/search_regression_baseline.json -> PASSED

Language Consistency

• rg -n "[\p{Han}]" content/cuda/docs | wc -l -> 0

Notes

• This PR updates and extends existing CUDA/PTX work already in branch history.
• PR Expand CUDA/PTX docs, improve retrieval aliases, and add search regression suite #175 is intentionally reused/updated rather than opening a new PR, to keep review context in one place.