chore(core,vm): drop the two dead extension includes and 29 stale externs (#744, items 1-2)

[Nitjsefnie]

What does this PR do?

Takes the two cheapest items of #744 — the dead extension includes in eigenscript.c and the stale extern block in vm.c — and leaves the header-split work alone. Declaration-only; the object files are byte-identical to the base.

Related Issues and Pull Requests

Relates to #744

Deliberately partial: this is items 1 and 2 of the issue's cheapest-fix-first ordering. The per-layer header split and the eigenscript.h umbrella are untouched, so #744 stays open.

Changes

• src/eigenscript.c: removed the guarded ext_db_internal.h and model_internal.h includes (:12-17). This drops the only core→ext include edge, and with it the core TU's libpq header dependency.
• src/vm.c: removed 29 of the 32 extern re-declarations (the block sits at :300-335 on current main, not the issue's :242-278 — Build: every variant writes src/eigenscript, which is why the test runner needs a binary fingerprint guard — a per-variant objdir retires both #740/build: per-variant objdirs — variants coexist, rebuilds go incremental (#740) #808 moved it), replacing them with a three-line block plus a comment saying why the survivors are there.

Correction to the issue's premise: it is 29, not 32. Three of them are load-bearing and stay — builtin_free_val, env_hash_find_dict and env_get_assign_count are called from vm.c (:2216, :3522, :5492, :510, :4784) and no header declares them. Giving them a header home is header work, so it belongs with the part of #744 this PR does not touch. Anyone reading the issue text alone will expect a 32-line deletion, hence flagging it here.

Of the 29 removed, 26 have a matching declaration in eigenscript.h (make_num :914, env_new :1050, observer_ensure_fresh :1157, val_type_name :1175, …) and 3 — val_incref, val_decref, num_guard — are static inline definitions there (:976, :984, :955), so the extern was inert and arguably ill-formed (one identifier with both internal and external linkage in a single TU). No type mismatch anywhere in the block: both declarations were visible in the same TU at the merge-base and gcc and clang accept it, so compatibility already held rather than being eyeballed. One of the removed comments claimed observer_ensure_fresh came from eval.c — a TU the bytecode VM replaced.

Testing

Merge-base cd9f823.

Behavior-neutrality, proven exactly rather than argued: I snapshotted build/{release,full,http,gfx}/{eigenscript,vm}.o from the unmodified merge-base, applied the change, rebuilt, and cmp'd — all 8 object pairs byte-identical, with both TUs confirmed to have genuinely recompiled. That is exact here because the release-family flags carry no -g and vm.c contains no __LINE__/__FILE__/assert(. The release suite log also diffs to zero lines against its pre-change baseline, and make freestanding-check's log is byte-identical.

Builds, commands lifted from .github/workflows/ci.yml: make, CC=gcc ./build.sh, CC=clang ./build.sh, make http, make gfx, make full, make asan — all pass. No new warnings; the clang warning multiset is identical to baseline once file:line is stripped (the vm.c 'len' may be used uninitialized line moves 4975 → 4950, exactly the 25 lines removed).

Suite (cd tests && bash run_all_tests.sh, direct exit code 0 on each): release 3364/3364, http+model 3468/3468, gfx 3405/3405, full 3504/3504, ASan+UBSan with detect_leaks=1 3368/3368 with no LeakSanitizer tally line.

The libpq consequence is measured, not inferred — compiling eigenscript.c alone with the full flags minus -I/usr/include/postgresql: before, fatal error: libpq-fe.h: No such file or directory; after, clean. (make full still passes that flag to every TU, since it is a variant flag in Makefile:74 — the win is that the core TU no longer needs it.)

On the includes, rather than sampling symbols I preprocessed two synthetic TUs — with and without the two headers — and diffed their identifier sets. The 511 identifiers the headers add intersect eigenscript.c in 18 places, all of which resolve to parameter/struct-member names inside declarations (PQexec's query, lo_read's buf, …) or #-line-marker filenames; a follow-up grep for member accesses on those names finds one hit, inside a comment. So zero declared entities are used, by whole-header enumeration rather than the issue's 7-symbol table.

Bugs Discovered

None filed — nothing found here is a defect in its own right.

Follow-ups / Known Limitations

• The three surviving externs want a header declaration; that is part of Layering is convention, not structure: no per-layer headers, eigenscript.h is a 1253-line umbrella, eigenscript.c carries two dead extension includes, vm.c has 32 stale externs #744's header work.
• Two related items left alone as out of scope: vm.c:624's extern Value g_null_singleton_external_decl; /* not used; doc only */ (I could find no definition for that symbol anywhere in src/), and the block-scope externs at vm.c:5177-5180 / :5221-5223, which are the issue's item 4.
• Pre-existing and unaffected: eigenscript.h:1288-1291 already declares ne_softmax_buf, ne_matmul_buf and json_obj_get, which model_internal.h:143-145,209 re-declares. That duplication is the header-split's problem, not this PR's.
• make net and make zlib were not built. Nothing in the diff is variant-conditional and those two differ from release only by a -D, so their objects for these TUs are identical to release's — happy to add them to the table if you want every variant represented.

Checklist

• make test passes locally — five variants plus ASan, all green, all at exit 0
• New builtins have signature comments and docs in docs/BUILTINS.md — n/a, no new builtins
• New library functions follow conventions in docs/STDLIB.md — n/a
• New examples have a comment header explaining what they demonstrate — n/a
• CHANGELOG.md updated (if user-facing change) — n/a, declaration-only with byte-identical objects; no user-visible change to record

Generated by Claude Opus 5 (brief, implementation, review)