Feature/new tiered int

[NikolayChirkov]

No description provided.

[NikolayChirkov]

Anton, please re-review when you have a chance.

I addressed the previous review comments and did an additional pass across the changed code, not only on the exact commented lines:

• kept the project at C++20;
• removed C++23-only usage;
• kept numeric core exception-free: no throw / stdexcept in numeric core;
• fixed constexpr validation so it works on GCC/MinGW without diagnostic pragma tricks;
• kept -fno-exceptions compatibility;
• restored required CI check names;
• kept MSVC on Visual Studio generator;
• removed duplicated workflow build/test steps where possible;
• applied kCamelCase for ordinary static constexpr members, while keeping STL-like value/type in traits;
• changed byte counts to std::size_t;
• replaced SFINAE-style stream checks with requires/concepts;
• added noexcept where safe;
• applied CamelCase function naming and east const in touched numeric code;
• updated assertions/preconditions where invalid use is a contract violation.

CI is green now.

I also rewrote the README manually after the API/documentation changes:

• clarified TieredInt serialization tiers and chunk sizes;
• fixed the wrong “first template parameter is the wire cell type” explanation;
• documented FixedPoint scale placement, including cases where the binary point is far outside the raw storage type;
• removed old local compiler paths from build docs;
• moved Exponential-over-TieredInt examples into Combined Types;
• added practical packed-size examples for payload sizes, latency exponential codes, and linear ping FixedPoint over TieredInt.

Please take another look.

[NikolayChirkov]

@BartolomeyKant, please re-review when you have a chance.

I addressed the latest comments and applied the same cleanup across analogous code:

• renamed the CMake target and public include root from numeric to ae-numeric, so public includes are now #include <ae-numeric/...>;
• updated install rules, README examples, tests, and the cpplint CI path accordingly;
• replaced DeserializeResult::BytesRead with bytes_read;
• rewrote the Exponential floating conversion through FloatingWorkCodec, FloatingToWork, and WorkToFloating, using fixed-width integer bit arithmetic;
• replaced the FixedPoint PromoteRep specialization table with generic integral promotion, including mixed signedness handling, while preserving same-rep support for packed reps;
• optimized PackedRing wrap copies with std::span plus memcpy and removed ordinal-only iterator comparison;
• kept the code C++20 and exception-free.

Local checks passed with MacPorts Clang 20.1.8 and CMAKE_OSX_DEPLOYMENT_TARGET=13.3:

• cmake --build build-mp-clang20-macos13.3-ae --config Debug --parallel;
• ctest --test-dir build-mp-clang20-macos13.3-ae --output-on-failure - 5/5 passed;
• cpplint --recursive ./ae-numeric.

GitHub Actions are green for fc29c6d: cpplint, Ubuntu GCC, macOS Apple-Clang, Windows MinGW, and Windows MSVC.

[BartolomeyKant]

Reread previous comments. Do not rewrite code just to pretend something is changed.
@NikolayChirkov your AI start hallucinating, validate what you're pushing.

[NikolayChirkov]

@BartolomeyKant I re-validated every open review thread against the current HEAD b82a367 and replied to each one individually with the actual current state.

Verified now:

• strict C++20; the former if consteval path is gone;
• GCC, Clang, and MSVC no-exceptions builds pass;
• current GitHub Actions run is green;
• fixed-width types are used for bit manipulation;
• FloatingWorkCodec<float> / FloatingWorkCodec<double> and generic conversion wrappers are present;
• mixed-signedness FixedPoint promotion is handled by numeric digit range, not sizeof alone;
• PackedRing uses std::span plus memcpy, and iterator equality uses pos_ rather than an ordinal;
• legacy TieredInt wire compatibility is covered by golden, exhaustive, randomized, and cross-decode tests;
• downstream aether-tele and aether-client-cpp migrations are green and preserve existing wire bytes.

I also called out the two remaining style nuances rather than claiming they were changed: the internal helper kWordPow is still misnamed as a function, and STL-style trait/type aliases intentionally remain lowercase. I have not resolved the review threads; please re-review the current diff when convenient.

[BartolomeyKant]

Issues in aether-tele should be fixed in this repository