core+lsp: one registration seam; the LSP builtin table is generated (#742)

[InauguralPhysicist]

What does this PR do?

Closes #742 — both halves: env composition goes through one registration seam, and the LSP's builtin table becomes generated, set-identical to the live registry.

Closes #742

Half 1 — the registration seam

• register_builtins now composes store (always compiled; freestanding gets ext_store's linkable no-op) and gfx (behind #if EIGENSCRIPT_EXT_GFX, matching the http/db/net/model rows) itself. The five hand-call sites — main.c ×2, eigs_embed.c, ext_http.c's worker-state path, lint.c ×3 (is_builtin_name, eigs_api_dump, E003) — are gone, and register_gfx_builtins gets a real declaration in eigenscript.h instead of main.c's private extern.
• The headline bug is fixed and pinned: a make gfx build used through the embedding API had no gfx builtins (gfx_open → undefined variable, reproduced red on pre-fix main with an embed probe against the gfx objects; green after). embed_smoke.c now asserts both compositions, and a new make embed-smoke-gfx target (extensions CI job) links the same smoke against build/gfx/*.o — headless, registration needs no SDL init.
• Band note: store/gfx move inside [0, g_builtin_binding_count), so eigs_is_registered_builtin now correctly claims them (sole consumer is build_corpus's identifier filter; previously it mis-classified store_open etc. as user names).

Half 2 — the generated LSP builtin table

• tools/gen_lsp_builtin_index.sh (same recipe and hygiene as gen_lsp_stdlib_index.sh: LC_ALL=C, zero dependencies, build artifact, never committed, .gitignored) generates src/lsp_builtin_index.h:
  • names from the registration seams (builtins.c, builtins_host.c, hash.c, ext_store.c) + every EIGS_*_BUILTINS group in ext_names.h — the LSP describes the language surface, not this binary's flags (the E003 binding-base rule);
  • hover detail from the /* name of ... */ signature comment above each C definition; the 124 builtins whose definitions never wrote one fall back to a BUILTINS.md pointer rather than papering the gap over.
• Result: 336 names (250 core + 86 extension), verified set-identical to --api --json's registry; 202 with signature detail — versus the hand table's 68 entries and phantom exec. Wired into the Makefile lsp target and build.sh lsp.
• The completed table surfaced the collision the issue's last paragraph predicted: mean is both a tensor builtin and stats.mean — the old table only "handled" it by not knowing the builtin existed (tests/test_lsp.py's define-shape-fallback check caught it immediately, 81/82). Hover now treats a dotted token as a member access: builtin/keyword/document-symbol lookups skip it and only the module-qualified stdlib branch may claim it. stats.mean hovers as the stats function; bare mean as the builtin.

Testing

• Release suite 3397/3397; ASan+UBSan with detect_leaks=1 3401/3401, zero sanitizer lines, leak floor 0 held.
• tools/freestanding_check.sh both stages green, allowlist untouched (store's registration under the profile is the existing no-op).
• LSP suite 82/82 (including the phantom-input exclusion and the repaired dotted-hover check); make lsp, make gfx, make http, make full all build.
• make embed-smoke and make embed-smoke-gfx green.
• Red-then-green on the embed+gfx bug: probe against pre-fix main's gfx objects fails undefined variable 'gfx_open'; this branch: OK.
• Generator planted-fault: registering a fake probe_gen_741x at the seam puts it in the regenerated table; the generator fails loudly on an empty scrape.

Follow-ups / Known Limitations

• 124 core builtins lack signature comments above their definitions (hover shows the BUILTINS.md fallback for them). Back-filling is mechanical, checklist-enforced for new builtins, and belongs in its own change — the generator prints the count on every run so it can't silently rot.
• Completion (as opposed to hover) still offers bare builtins after a module. prefix — pre-existing behavior with the old table too; noting rather than expanding scope.

Checklist

• make test passes locally — 3397/3397 release, 3401/3401 ASan+UBSan leak-clean
• New builtins have signature comments and docs in docs/BUILTINS.md — n/a, no new builtins
• New library functions — n/a
• New examples — n/a
• CHANGELOG.md updated

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR addresses #742 by centralizing global environment builtin registration behind a single register_builtins seam (fixing missing gfx builtins in embed+gfx builds), and by generating the LSP builtin hover/completion table from the live registration sources so it can’t drift.

Changes:

• Move gfx (when built) and store (always) registration into register_builtins, removing per-entry-point hand registration.
• Add tools/gen_lsp_builtin_index.sh and wire it into make lsp / build.sh lsp to generate src/lsp_builtin_index.h (and ignore it in git).
• Update LSP hover logic to treat dotted tokens (x.name) as member access to avoid builtin-vs-stdlib collisions (e.g. mean vs stats.mean), and add CI coverage for embed+gfx (make embed-smoke-gfx).

Reviewed changes

Copilot reviewed 13 out of 14 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tools/gen_lsp_builtin_index.sh	New generator for LSP builtin docs table (names + signature-comment scrape).
src/builtins.c	register_builtins now composes gfx/store, making it the single env composition seam.
src/main.c	Removes hand registration of gfx/store; relies on register_builtins composition.
src/eigs_embed.c	Embed runtime init now uses only register_builtins (seam consolidation).
src/ext_http.c	Worker thread env init now uses only register_builtins (seam consolidation).
src/lint.c	Removes explicit store registration; uses the composed seam env consistently.
src/eigenscript.h	Adds public prototype for register_gfx_builtins.
src/embed_smoke.c	Adds assertions that store (and gfx when built) are present via embed seam.
src/eigenlsp.c	Replaces hand builtin table with generated header; updates hover for dotted tokens.
Makefile	Adds lsp builtin index generation and a new embed-smoke-gfx target.
build.sh	Ensures builtin index header is generated in the lsp build path.
.gitignore	Ignores generated src/lsp_builtin_index.h.
.github/workflows/ci.yml	Runs make embed-smoke-gfx in CI.
CHANGELOG.md	Documents the seam fix and the generated LSP builtin index.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 13 out of 14 changed files in this pull request and generated no new comments.

Suppressed comments (2)

Makefile:220

• The lsp_builtin_index.h generator scrapes signature comments from additional TUs (builtins_tensor.c, ext_gfx.c, ext_http.c, ext_db.c, ext_net.c, model_train.c), but the Makefile rule only depends on the seam sources + ext_names.h. That means changing a builtin’s signature comment (or moving/adding one) in those files won’t regenerate the header on incremental builds, leaving eigenlsp hover text stale until a clean build.

# Builtin half of the same idea (#742): names from the registration seams +
# ext_names.h, hover detail from the signature comments. Also a build
# artifact, never committed.
$(SRC_DIR)/lsp_builtin_index.h: $(SRC_DIR)/builtins.c $(SRC_DIR)/builtins_host.c \
		$(SRC_DIR)/hash.c $(SRC_DIR)/ext_store.c $(SRC_DIR)/ext_names.h \
		tools/gen_lsp_builtin_index.sh
	bash tools/gen_lsp_builtin_index.sh

tools/gen_lsp_builtin_index.sh:35

• The script comment says it exits nonzero "and no output written" on parse failure, but currently pipeline failures (e.g., missing source files, grep/sed/awk errors) won’t necessarily stop the script before the final mv, so a partial/empty header could be installed. Enabling -e + pipefail makes failures loud and prevents writing a broken lsp_builtin_index.h.

set -u
cd "$(dirname "$0")/.."