Mark exported classes in a static block and identify by the mark alone

[backspace]

Stacked on #6614. This PR applies the review feedback on #6614, so the two can be compared before one is chosen.

Background and Goal

#6614 marks each exported class or function with the module that declares it. A helper function is appended to every module, and the loader still keeps its own identity map. Review feedback asked for less: mark only exported classes, make Loader.identify read only the mark, use Babel's visitor pattern, and use the host's existing Babel config instead of the bundled-base Vite plugin. This PR makes those changes.

Where to start

• moduleProvenancePlugin in runtime-common/loader-plugin.ts. It visits ExportNamedDeclaration and ExportDefaultDeclaration, and puts a static block first in each exported class that the module declares.
• Loader in runtime-common/loader.ts. identify is now moduleProvenanceOf(value). markModuleExports replaces captureIdentitiesOfModuleExports.
• host/babel.config.cjs. An overrides entry applies the plugin to packages/base files, with a literal @cardstack/base/<name> identifier.

Feedback, and what this PR does

Feedback	Change
identify should only use the provenance mark	Done. Loader.identify returns moduleProvenanceOf(value). The instance identify method and the per-loader identity map are removed.
Use the existing Babel config, not the Vite plugin	Done. bundled-base-scoped-css.mjs handles only scoped CSS again.
Use ExportNamedDeclaration in the visitor	Done. There is no Program exit pass.
Marks too much: only exported classes matter	Done. Functions, constants and other values get no mark. All identify callers pass card or field classes.
The inlined helper adds too much code; define a global	Not adopted. Each class gets one static block (one Object.defineProperty call). There is no helper, so a module with no exported class gets nothing. A global would make every marked module depend on code that must evaluate before it, which is the module-ordering risk seen in CS-13655.
captureIdentitiesOfModuleExports should not be needed	Reduced, not removed. See below.

Key decisions and non-obvious mechanics

• The mark has to be in canonical form. A code ref stores the virtual-network form of a module identifier (for example @cardstack/base/card-api), not the URL. identify no longer has a loader to convert a URL, so the loader adds import.meta.moduleIdentifier (already in canonical form) to each module it evaluates, and the mark reads it. A module evaluated without a loader falls back to import.meta.url.

• Why the loader still marks exports. markModuleExports does two jobs:

  • It records Loader.loaders for Loader.getLoaderFor, which card-api, operations and patch-fields still use.
  • It marks exports that have no mark yet: classes in a module the loader gets as a shim (many host tests shim card classes under realm URLs), and classes in a module transpiled before marks existed. The L2 transpile cache has no transpiler version in its key, so unmarked output can be served for some time after a deploy.

  A value that already has a mark keeps it. So the mark from transpilation always wins, and identify has one source.

• The mark is in a static block, first in the class body. So the mark is set when the class is created, before any other static code runs. That includes cases where a circular import reads the class before its module finishes evaluating. A subclass doesn't run its parent's static block, and moduleProvenanceOf reads only an own property, so a subclass never takes its parent's mark.

• One behaviour change for shims. Before, each loader kept its own identity map. Now a shimmed class gets one mark for all loaders, and the first shim wins. Production shims (base, externals, host tools) always use the same identifier, so only a test that shims one class object under two different URLs would see a difference.

🤖 Generated with Claude Code

https://claude.ai/code/session_015idm3FXE4bujPmH1obGWFG

[github-actions]

Preview deployments

• Host staging preview

• Host production preview

Host Test Results

    1 files      1 suites   1h 9m 30s ⏱️
5 209 tests 5 198 ✅ 11 💤 0 ❌
5 224 runs  5 213 ✅ 11 💤 0 ❌

Results for commit fb02357.

Realm Server Test Results

    1 files    319 suites   2h 2m 1s ⏱️
5 104 tests 5 104 ✅ 0 💤 0 ❌
5 163 runs  5 163 ✅ 0 💤 0 ❌

Results for commit fb02357.