RV32: active data-segment initializers are silently dropped — ship them (linker-script placement + startup copy) and hard-decline until then

[avrabe]

Found by the VCR-VER-003 phase-2 probe (#777, PR #797)

The RV32 backend uses a single-base scheme (s11 = __linear_memory_base, provided by the generated linker script at _bss_end, zeroed at reset). The relocatable object ships .text only — neither the object, the generated startup.c, nor linker.ld carries the wasm module's active data segments. A module like

(module
  (memory 1)
  (data (i32.const 16) "\01\02\03\04")
  (func (export "get") (result i32) i32.const 16 i32.load))

compiles clean, and get returns 0 at runtime instead of 0x04030201 — the #757-family silent drop (worse: unconditional, not overlap-dependent).

What phase 2 landed (PR #797)

validate_served_image(segments, &[]) runs unconditionally on the RV32 emit path and WARNS LOUDLY per nonzero un-served initializer byte (later-wins semantics: all-zero / zero-overwritten runtime images are genuinely served by zeroed RAM and stay silent).

Why it is a warning, not a hard decline (yet)

The CI-pinned RV32 frozen fixture (control_step.wasm, frozen_fixtures_rv32_text_is_bit_identical_oracle_001) carries a nonzero active segment (11 B at 65536) that its differential never reads — a hard decline would break the frozen compile-success gate. Flipping to a decline needs the refreeze ritual (or the real fix below) in its own evidence-gated PR.

The real fix (RV32 analogue of #758)

1. Emit the packed initializer image into the object (e.g. .wasm_data PROGBITS, later-wins dense image or per-segment with offsets),
2. place it via the generated linker script at __linear_memory_base + off (load address in flash),
3. copy ROM→RAM in the generated startup (the existing _data_load/_data_start pattern),
4. validate the shipped image with the existing validate_served_image (turn the warning into the mixed-split-style hard error on mismatch),
5. differential: an RV32 module that READS its data segment vs wasmtime.

Until then the honest end-state per the #615 expand-or-loud-reject rule is a hard decline for nonzero-initializer modules on -b riscv.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L