Reject Bedrock2 C outputs built for the wrong word size

[JasonGross]

Summary

• emit an exact target-word-width assertion in the Bedrock2 C prelude
• regenerate all checked-in Bedrock2 C outputs with the assertion
• compile only native-width outputs in the generated-C test and verify that a mismatched output is rejected
• document the Bedrock2 target-width requirement

Rationale

Bedrock2 uses uintptr_t for both addresses and arithmetic words. Previously, compiling a _32.c output for a 64-bit target made _br_load and _br_store access eight bytes while generated limb offsets advanced by four bytes. That caused overlapping limb accesses, a four-byte overrun at the end of caller buffers, and incorrect field arithmetic.

The generator already receives the synthesis word size. This change preserves Bedrock2’s pointer-sized word model and turns any mismatch into a compile-time error before the affected code can run. The checked-in generated artifacts now reject both 32-on-64 and 64-on-32 mismatches. This addresses Scrutineer finding #2509.

Testing

• opam exec -- make src/Bedrock/Field/Stringification/Stringification.vo EXTERNAL_DEPENDENCIES=1 -j2
• make -f Makefile.examples only-test-bedrock2-files CC=cc
• make -f Makefile.examples only-test-bedrock2-files CC=clang
• git diff --check

A genuine 32-bit compiler/sysroot was not available locally, so successful _32.c compilation on a 32-bit host was not executed. The generated assertion is symmetric, and its rejection path is covered on this 64-bit host.

Authorship note: this was researched and written by an AI coding agent
(OpenAI Codex), working on Jason Gross's behalf; Jason reviews what is
posted from this account.

[JasonGross]

The five platform-specific failures had the same cause: the checked-in Bedrock2 outputs contained one extra blank line before the new assertion compared with freshly generated output. The 31 artifacts now match the generator. GCC and Clang both compile the native-width corpus and reject the mismatched-width sample with the expected diagnostic. The Alpine failure was unrelated infrastructure trouble while setup-alpine downloaded apk-tools.

Authorship note: this was researched and written by an AI coding agent
(OpenAI Codex), working on Jason Gross's behalf; Jason reviews what is
posted from this account.