core: move BRIDGE_REGISTER_MODEL/BRIDGE_REGISTER_ACTION's registrar out of the header

[Yaraslaut]

Summary

BRIDGE_REGISTER_MODEL/BRIDGE_REGISTER_ACTION's registrar initialiser is an ordinary function call written where the macro is expanded, so its body — and everything it drags in (the action's glaze codecs, the dispatcher's runner closure, forms::schemaJson<A>'s schema generator) — is instantiated in every translation unit that includes a model header carrying it. A prior measurement on examples/kanban/include/kanban/models/board_model.hpp found 26.93 CPU-seconds of -O3 codegen per including TU, for codecs the overwhelming majority of those TUs never call.

• BRIDGE_DECLARE_MODEL/BRIDGE_DECLARE_ACTION decompose the existing macros into a header-safe half (the ModelTraits<M>/ActionTraits<A> specialisation alone, free to repeat per TU) and BRIDGE_REGISTER_MODEL_SOURCE/BRIDGE_REGISTER_ACTION_SOURCE carry the registration itself, called once in the .cpp that should own it. The combined BRIDGE_REGISTER_MODEL/BRIDGE_REGISTER_ACTION are unchanged and remain the right choice when the split isn't worth it.
• Splitting declaration from registration introduces a mistake that couldn't previously happen: a model declared but never registered anywhere. Closed with a link-time canary — BRIDGE_DECLARE_MODEL/BRIDGE_DECLARE_ACTION call a function template behind extern template, and only the SOURCE macros explicitly instantiate it — so a missing registration fails to link, instead of compiling clean and surfacing only as a runtime "unknown model type" far from the cause.
• Retrofits examples/kanban's board_model.hpp/.cpp — the exact header the original measurement used — as proof this reaches the measured cost. Re-measured (AppleClang 17, not isolated from concurrent system load on this run, so reported as a lower bound): 20.81 CPU-s before, 7.81 CPU-s after, a 13.00 CPU-s reduction per translation unit; the compiled stub's object file shrinks from 3,331,704 to 3,232 bytes (7155 to 24 symbols), with zero glaze symbols left in the post-fix object.
• docs/spec/core/registry.md gains "Moving a registrar out of the header", documenting the mechanism, the canary, and its real gaps (a dlopened plugin module must keep using the combined macros; the static-library archive-drop interaction is stated as inferred, not measured).

Closes #791

Test plan

• Full core suite (morph_tests, 1602 test cases / 23513 assertions) green, both before and after a rebase onto current master.
• Full examples/kanban ladder suite (ladder_kanban_tests, 172 test cases / 1192 assertions) green.
• New tests/compile_checks/declare_only_no_source_link.cpp + declare_only_source_provided.cpp, wired through a new try_compile() pair in tests/CMakeLists.txt, prove the canary both directions: a declared-but-unregistered model/action fails to link, and the same program links cleanly once the SOURCE macros are added.
• examples/kanban's ladder_kanban_lib, ladder_kanban_server, and ladder_kanban_headless all link successfully against the retrofitted board_model.hpp/.cpp.
• Before/after CPU-second measurement on board_model.hpp (methodology matching the original issue: -fsyntax-only and -O3 -c, best of two, on a stub TU that only includes the header), corroborated by object-file size and symbol-count diffs.

🤖 Generated with Claude Code

https://claude.ai/code/session_018fEUahMFF32wQLiWjbsfkc

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!