build: carry -Werror=switch in build.sh and the sanitizer/valgrind legs

[Nitjsefnie]

What does this PR do?

The follow-up you okayed on #779: ./build.sh and the asan|asan-http|tsan|valgrind legs now carry -Werror=switch, so the gate bites on the builds a contributor actually runs locally rather than only in CI via make http|gfx|zlib|lsp|full.

The line numbers in #779's note had drifted, and there were more lists than it said. Makefile:184/206/219/233 are the target-name lines; the recipes are one line lower in two cases. And build.sh has three flag lists, not one. Actual edits:

file:line	leg	change
Makefile:185	asan	after -fsanitize=address,undefined
Makefile:206	asan-http	after -fsanitize=address,undefined
Makefile:220	tsan	after -fsanitize=thread
Makefile:234	valgrind	before -g -O1
build.sh:45 · :55 · :67	lsp · full · minimal	after -Werror=implicit-function-declaration

Flag ordering follows Makefile:9's CFLAGS. No shared variable was factored out — the only shared list is $(CFLAGS) and these seven deliberately don't use it, so leaving them alone keeps the diff to the flag itself. No other warning flag touched.

Checklist

• make test passes locally — RESULTS: 3275/3275 passed, 0 failed
• New builtins have signature comments and docs in docs/BUILTINS.md — n/a, build-flag change
• 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, no user-visible behaviour change; the compiler gate is a contributor-facing build detail

The evidence that matters is that every leg still compiles with the gate armed, since that is the whole risk of the change:

• CC=gcc ./build.sh → EigenScript 0.34.0 built. Binary: 764K; CC=clang ./build.sh → 788K
• make asan → (asan+ubsan) built · make asan-http → (asan+ubsan, http+model) built · make tsan → (tsan) built
• Regression check on the legs that already carried it: make http → 880K, make gfx → 808K, make zlib → 772K, make lsp → 924K, make full → 880K
• Zero switch diagnostics anywhere. Nothing was non-exhaustive on a path CI had not been checking, which is the reassuring outcome but not one I would have asserted without building each leg.

One leg I could not verify locally, stated plainly: make valgrind does not build on this box — src/eigenscript.c:29:10: fatal error: valgrind/memcheck.h: No such file or directory, because valgrind is not installed and I did not install system packages. That is a missing toolchain, not a switch failure: compilation dies at preprocessing before any switch is seen. Your ci.yml:455-457 installs valgrind, so that leg gets exercised there. The -DEIGS_VALGRIND code behind it is memcheck macro calls rather than switch logic, so I would be surprised if it bites — but I have not proven that and am not claiming it.

Follow-ups / Known Limitations

Six more hand-written flag lists still omit -Werror=switch: Makefile:148 (jit-smoke), :251 (poison), :301 (coverage), :322/:339 (the fuzz harnesses), :364 (freestanding). They are outside what you approved, so I left them — happy to send them as a second pass if you want the whole perimeter.

Unrelated pre-existing warnings noticed while building, not introduced here and not failing anything: lint.c:813/:832 -Wcomment (/* within a comment), lint.c:899 -Wformat-truncation, and under clang step.c:491 -Wunused-but-set-variable on int fed = 0;.

Generated by Claude Opus 5 (brief, review), Kimi K3 (implementation)

[InauguralPhysicist]

Review — approved

This is the follow-up agreed in #779, and it does exactly what it says: -Werror=switch added to the seven flag lists a contributor actually hits locally (build.sh ×3, asan, asan-http, tsan, valgrind) and nothing else. I diffed each hunk against the Makefile/build.sh on main — flag placement follows CFLAGS ordering, no other flag touched, no factoring detour.

Points in your favor:

• Correcting my line numbers. You're right that #779's note pointed at target-name lines and undercounted build.sh's flag lists — three, not one. Good catch rather than silently following a wrong map.
• Building every leg both compilers rather than asserting from the diff that nothing could break. "Zero switch diagnostics" is the load-bearing claim of this PR and you proved it per-leg instead of inferring it.
• The valgrind caveat, stated plainly. Correct call not to install system packages. CI settles it: the valgrind (memcheck) leg ran on this PR and passed, so the one gap in your local sweep is now closed with evidence.

On the follow-ups:

• The six remaining lists (jit-smoke, poison, coverage, fuzz ×2, freestanding) — yes, send the second pass. Same shape as this one is fine. The fuzz harnesses in particular compile subsets of $(SOURCES), so they're the most likely place a non-exhaustive switch could hide unchecked.
• The pre-existing warnings you noted (lint.c -Wcomment/-Wformat-truncation, step.c unused-but-set under clang) — worth an issue so they don't get lost; they're real even if harmless today.

CHANGELOG n/a is the right call — contributor-facing build detail, matches this repo's convention.

CI: all 16 checks green. Merging.