Version-derived feature flags in the node implementor misfire on runtimes that report a fictional Node version

Commencez par implementors/node/features.js et le README du node implementor ; examinez les vérifications existantes de process.version ainsi que le précédent dans #68. Décidez si le node implementor doit documenter son hypothèse d’exécution dans un Node réel ou détecter les capacités avant d’activer les flags. C’est terminé lorsque le comportement choisi empêche des runtimes étrangers de tenter d’utiliser des cas SharedArrayBuffer non pris en charge.

implementors/node/features.js derives experimentalFeatures.sharedArrayBuffer and
runtimeFeatures.dataviewSharedArrayBuffer from process.version. While exploring how
far foreign runtimes get with the suite, I ran it through the node implementor under
Deno 2.9.5, and its Node compatibility mode reports process.version as v26.3.0, a
Node version that does not exist. The version comparison therefore enables the
SharedArrayBuffer cases, but Deno exports neither node_api_create_sharedarraybuffer
nor node_api_is_sharedarraybuffer, so the addon fails at load time instead of the test
being skipped: on Linux the process exits 127 on the unresolved symbols, and on Windows
dlopen reports it cannot load the addon. The suite output then reads like a broken
test rather than a missing feature. Bun does the same kind of impersonation, its stable
1.3.14 reports v24.3.0 and its current canary also reports v26.3.0, so any invented
version string can end up in these comparisons.

For Node itself the version checks are exactly right, and #26 already made features.js
a per-implementor file, so an eventual bun or deno implementor would declare its own
honest flags. The situation above only exists in the period before such implementors do,
which is also exactly when someone evaluates a new runtime by pointing the node
implementor at it. So a small question rather than a proposal: is that simply out of
scope for the node implementor (worth a sentence in its README that its flags are only
meaningful under real Node), or would you want flags like these derived from a probe of
the actual runtime rather than from the version string? #68 adds one more flag in the
current version-derived style, so whatever you prefer would apply there too. Happy to
help with either outcome.