docs: enhance AGENTS with project context and performance guardrails

[napetrov]

Summary

Follow-up to #276 — adds Intel-specific technical detail and performance guardrails based on code review.

Changes

Root AGENTS.md

• Add project description, tech stack (C++20, MKL, OpenMP, pybind11)
• Add core principle: performance over simplicity in hot paths
• Add quick start commands (build/test/Python)
• Add directory map for navigation

.github/copilot-instructions.md

• Add concrete authoring rules (e.g. don't refactor full files for 2-line changes)
• Add contribution expectations: backward compat, pre-commit, regression tests
• Expand source-of-truth: explicitly list mkl.cmake, multi-arch.cmake, numa.cmake, pyproject.toml
• Add negative constraint: no std::iostream in perf-critical headers

include/svs/AGENTS.md

• Add perf constraints: memory alignment, SIMD dispatch rules, hot path rules
• Add failure modes: template instantiation impact, ABI compat

bindings/python/AGENTS.md

• Add pybind11 rules: GIL release, dtype/shape validation, memory ownership
• Add common failure modes: wrapper sync on C++ signature changes, edge cases, lifetime management

cmake/AGENTS.md

• Document Intel-specific modules: mkl.cmake (threading), multi-arch.cmake (ISA dispatch), numa.cmake, openmp.cmake
• Add guardrail: no removal of optimization flags without benchmark validation

benchmark/AGENTS.md

• Add constraints: warmup/iteration stability, dataset provenance, baseline comparisons
• Add reporting requirements: median/min/max/stddev

include/svs/quantization/AGENTS.md (new)

• Quantization guardrails: type safety, precision bounds, SIMD alignment, codec compat
• Common failure modes: float arithmetic in hot paths, misaligned access, codebook sync

Notes

• No mutable details (versions, flags) hardcoded in instruction files
• All changes are additive — no policy removed from original PR

[mihaic]

Looks good. Please remove LVQ references according to suggestions.

[mihaic]

No LVQ in this repo.

Suggested change

	High-performance C++ library for vector similarity search at billion scale. Uses Intel MKL, AVX-512/multi-arch SIMD dispatch, LVQ quantization, NUMA-aware memory, OpenMP threading. Python bindings via pybind11. Archetype: **C++** (with Python bindings).
	High-performance C++ library for vector similarity search at billion scale. Uses Intel MKL, AVX-512/multi-arch SIMD dispatch, quantization, NUMA-aware memory, OpenMP threading. Python bindings via pybind11. Archetype: **C++** (with Python bindings).

[mihaic]

Suggested change

	Quantization algorithms and data types (LVQ, scalar quantization).
	Quantization algorithms and data types (scalar quantization).