only-include-used-components: both hand-written maps can drift silently

[kevbarns]

Follow-up to #505. only-include-used-components relies on two hand-written maps in src/bin/only-include-used-components.ts, and neither has a guard. Both failure modes are silent: the script still exits 0.

1. REACT_DSFR_MODULE_TO_DSFR_COMPONENTS has no exhaustiveness check

A module absent from it and from NON_COMPONENT_MODULE_IDS makes resolveModuleIdToDsfrComponents return undefined, which trips the include-everything fail-safe. Adding a component to src/ is enough to disable the optimization for every consumer importing it.

Reproduced on #512: resolveModuleIdToDsfrComponents({ moduleId: "ContentMedia" }) returns undefined. Same result for the directory form (src/ContentMedia/index.tsx) and for a new src/blocks/* entry.

The maintainer merging such a PR gets no signal, and no existing test fails.

2. content is detected on core utility classes, causing over-inclusion

DSFR_COMPONENT_DETECTION_CLASS_PREFIXES["content"] lists fr-responsive-img and fr-responsive-vid. Those are core classes, not content classes:

• .fr-responsive-img and its 7 ratio modifiers are defined in core.main.css, which is always included.
• All 8 occurrences in content.main.css are scoped under .fr-content-media, 0 unscoped.

So any project using fr-responsive-img pulls in the whole content component for nothing. Measured end to end on a consumer using fr-responsive-img without fr-content-media:

	components	dsfr.min.css
current	3 / 45	234 877 B
without the two prefixes	2 / 45	231 924 B

Dropping them is safe: the only content rules involving those classes require .fr-content-media, which is already a detection prefix on its own.

Proposal

Two tests that re-derive each map from a machine source rather than restating it, plus the one-line fix for 2:

1. Enumerate the public subpaths from the real src/ layout (minus what src/tsconfig.json excludes), run them through the production module id extraction, assert each resolves.
2. Assert every detection prefix opens a selector in its own component stylesheet, read from the installed @gouvfr/dsfr. Merely appearing in it is not enough, which is exactly what 2 gets wrong.

Both should be checked against the case that makes them answer false, not only the nominal one.

I have this ready and will open a PR.