Env composition is duplicated across 5 call sites: `make gfx` + the embedding API yields a runtime with no gfx builtins, and eigenlsp's builtin table is 24% complete

[InauguralPhysicist]

Found in a modularity review of main @ 8e8970a. Tracking issue. The registration seam is clean for 3 of 5 extensions; the other two bypass it, and the "compose the global env" sequence is open-coded in several places that have already diverged.

1. Two extensions bypass the single registration seam

builtins.c:7386-7396 registers http/db/model behind exactly three #ifs — a genuine single point. But gfx and store are registered by hand elsewhere:

eigs_embed.c:25    register_store_builtins(global);
ext_http.c:1280    register_store_builtins(global);
lint.c:152         register_store_builtins(e);
lint.c:2385        register_store_builtins(e.bind);
main.c             register_store_builtins(...) / register_gfx_builtins(...)

register_gfx_builtins is declared at main.c:11 and called only from main.c. So a make gfx build used through the embedding API has no gfx builtins at all — eigs_embed.c never calls the gfx registrar. EIGS_GFX_BUILTINS already exists in ext_names.h:27 precisely to make that cheap, so the mechanism is present and unused.

A fifth registrar today requires N hand-edits across these sites: the exact drift class ext_names.h was built to eliminate, unsolved one level up.

Fix: one register_all_builtins(Env*) that every entry point calls, with the #ifs inside it.

2. Four builtin-name registries, only two gated — and the ungated one is 24% complete

235 builtins are registered in builtins.c. Four independent registries must track them:

Registry	Gated?	State
SANDBOX_ALLOW[] (builtins.c:4271-4320)	yes — tests/test_sandbox_allow.eigs	148 entries, 0 phantom
docs/BUILTINS.md / SPEC.md / STDLIB.md	yes — tools/stdlib_index_check.sh	clean
builtin_docs[][2] (src/eigenlsp.c:26)	no	68 entries — 167 of 235 builtins absent, plus a phantom exec
lsp_stdlib_index.h	generated	correct by construction

Verified by counting: awk '/builtin_docs\[\]\[2\]/,/^};/' src/eigenlsp.c | grep -c '{"' → 68.

The two gated registries are at zero drift. The ungated one fell to 24% coverage precisely because nothing checks it — so LSP hover and completion silently omit three quarters of the language. This is the cleanest natural experiment in the repo for the value of a gate, and the fix is to point tools/stdlib_index_check.sh's existing extraction at eigenlsp.c too (or generate builtin_docs the way lsp_stdlib_index.h is already generated).

Related: map/filter/reduce/reverse appear in both the hand-written table and the generated stdlib_docs, with no reconciliation between them.