C++ embedding migration: current targets

The broad #include <kyna/kyna.hpp> umbrella was removed. Include and link the smallest public domain interface needed by the embedding:

• full checking/running: kyna/language/language_session.hpp, link kyna_embedding;
• token inspection: kyna/lexing/tokenizer.hpp, link kyna_lexing;
• parsing: kyna/parsing/module_parser.hpp, link kyna_parsing;
• module loading: kyna/modules/module_loader.hpp, link kyna_resolution;
• analysis: kyna/semantics/module_analyzer.hpp, link kyna_typecheck;
• execution: kyna/execution/bytecode_virtual_machine.hpp, link kyna_vm.

Use tokenize(SourceFile), parseModule(SourceFile, tokens), and analyzeModuleGraph(graph) result objects instead of catching exceptions for expected source mistakes. LanguageSession is the normal high-level replacement for manually sequencing the v0.1 lexer, parser, analyzer, and interpreter.

Host access is injected through RuntimeCapabilities. Production adapters are the default; deterministic tests can provide filesystem, process, network, and clock ports. The current build does not declare the old kyna_lib aggregate; use the owning target above. The public tree-walk headers remain available through kyna_vm during migration, but their syntax/environment contracts cannot yet be replaced by VM-backed adapters.

LanguageSessionOptions.collectMetrics opts into phase timing in LanguageResult.metrics. Existing native adapters implementing BytecodeNativeAdapter::invoke continue to work. Adapters needing language callbacks can override invokeWithCallbacks; its callback interface is borrowed for that invocation only. Root managed values across re-entry. See the milestone inventory for remaining compatibility work.