feat: implement Kompact v1 foundation

[trancee]

Summary

• implement the Kompact v1 common runtime, annotations, KSP2 processor, Gradle plugin, generated Kotlin/C99 interfaces, registry compatibility, and publication model
• validate reserved/unknown identities, versions, lengths, transport-tail bits, and recursively ordered bodies with validate-before-mutate behavior
• generate scalar, byte, array, optional, nested, nested-array, active, and decode-only Kotlin/C interfaces from one canonical descriptor pass
• add stable schema diagnostics, cycle and wrapped-width checks, canonical rational semantics, duplicate-key rejection, lifecycle/history/source validation, descriptor fingerprints, and ABI baselines
• make standalone KSP work with separately published source-retained annotations; validate aligned runtime/annotation/plugin versions
• add deterministic relocated-cache tests, exact generated C snapshot checks, disposable JVM/Android/iOS/C publication consumers, and atomic output publication
• add reviewed conformance cases, fixed-seed Kotlin property coverage, C status/precedence and sanitizer fuzz harnesses
• add retained small/medium/large workloads, optimized byte-chunk runtime operations, JVM ratio enforcement, AndroidX Microbenchmark, and Apple benchmark compilation
• add Linux, Android, Apple, QEMU big-endian, fuzz, publication, performance, and generic physical-runner workflows

Specification

Implements the repository-controlled requirements indexed by the closed Kompact v1 implementation-ready specification and ADR-0001 through ADR-0010.

TDD evidence

Observed red then green at generated-consumer seams for transport-tail rejection, validation precedence, C aggregate writers/optionals, optional arrays, schema cycles, array-count diagnostics, semantic names, lifecycle transitions, namespace mismatch, duplicate registry keys, supported-decoder removal, published annotation discovery, and real multi-target publication consumption. Runtime property and benchmark-workload correctness tests cover the non-generated seams.

Verification

• ./gradlew clean --no-build-cache --no-configuration-cache && ./gradlew check --no-build-cache --no-configuration-cache passed: 163 actionable check tasks, 157 executed
• disposable repository publication passed: 123 actionable tasks; the external consumer then compiled generated code for JVM, Android, iosArm64, and iosSimulatorArm64
• strict multi-translation-unit generated C99 consumer passed with -Wall -Wextra -Wconversion -Wsign-conversion -Werror -pedantic-errors
• generated large C header: 17,111 bytes, below the 64 KiB limit
• full JVM benchmark profile passed the 1.10 reference ratio gate: large 0.9635, medium 0.9863, small 0.9724
• AndroidX benchmark APK assembled; iOS benchmark sources compiled for device and simulator
• GCC C conformance, fixed-seed C fuzz smoke, Clang ASan/UBSan fuzz execution, ABI checks, Spotless, detekt, Kover, Android lint, and benchmark smoke passed
• actionlint and git diff --check passed

Review

The two-axis /code-review pass produced one valid Standards finding: the 300-line D9 default deviations lacked a repository rationale. README.md now records each cohesive transaction/fixture deviation; every maintained file remains below the 500-line hard ceiling. The current-index Spec recheck found no remaining decode-only or C enum-array issue and reported no findings.

External release blockers

• the maintainer selected generic android-reference and ios-reference self-hosted labels because concrete runner/device identities are unavailable; replace them with pinned metadata before release
• the Android physical workflow is executable; the iPhone workflow currently captures metadata and compiles iosArm64, but signed device execution, timing loops, and Instruments allocation evidence still require the physical runner/signing environment
• no production firmware compiler is selected; GCC and Clang remain the defined gates until a firmware toolchain is adopted
• obtain independent human review

The PR remains draft until these external gates complete.

Impact

This remains a first pre-release API and wire-format implementation; no published consumer migration exists. Runtime failures remain typed and redacted and never include packet or attempted write values. The runtime bit loop now processes byte chunks rather than individual bits; controlled JVM generated/reference ratios remain within the ADR-0008 ceiling.