fix(globals): #649 i64.const global initializers reach the emitted image — decoder captures both words, startup materializes the R9 table

[avrabe]

Fixes #649. Lineage: #645 (i64 global get/set pair lowering + width-aware slot layout), #643, #648/#369 (GI-FPU-001 float loud-skip lane).

Mechanism

WasmGlobal.init_i32 only decoded a leading i32.const; an i64.const init decoded to None and every consumer's unwrap_or(0) silently ZEROED the slot. #645 made the accesses width-correct but a nonzero i64 init still never reached the output — worse, the default self-contained image never materialized a globals table at all (R9 was never initialized; reads hit whatever R9=0 pointed at).

Fix

Decoder (synth-core/wasm_decoder.rs): init_i32: Option<i32> → init: Option<GlobalInit> with I32(i32)/I64(i64) — a rename, so the compiler finds every consumer (no wildcard survives). f32/f64 inits deliberately stay None: float global access is the GI-FPU-001 loud-skip lane (#648) and capturing bit patterns here must not quietly unskip it.

Consumers audited (all of them, forced by the rename):

consumer	handling
identify_stack_pointer_global (#237 SP promotion)	matches GlobalInit::I32 only — an SP is an i32 address; i64 inits are data, never promoted
--native-pointer-abi __synth_globals slots (NativeGlobalsLayout, all data/bss write sites in build_relocatable_elf)	stays 4-byte i32 slots; wide globals already REFUSED up front (#643 bail = the decline-loudly arm), so only I32 inits can land; byte-identical for i32-global modules vs origin/main
self-contained Cortex-M image (build_multi_func_cortex_m_elf + single-function build_cortex_m_elf)	previously NO materialization anywhere: now generate_minimal_startup receives the table words (#643 summed-width layout, i64 = two words lo-first), sets R9 = 0x2000_0000 + linmem_size, and stores every captured init word (via the R12 scratch — reserved, and no user code has run yet) before BLX; RAM sizing + the NoBits phdr/section cover the table; tables past the STR.W #imm12 range decline loudly
plain --relocatable (host-owned R9 table)	unchanged contract — the host provides the table (the #643 oracle pins it)
RV32	no global lowering: every global access already loud-skips (pinned by the #643 oracle)

No globals ⇒ the historical 28-byte startup blob and the whole ELF stay byte-identical (cmp vs an origin/main build: identical; frozen anchors 10/10).

Red → green

scripts/repro/i64_global_init_649_differential.py + i64_global_init_649.wat — the DEFAULT cortex-m4f image (the exact #649 verification shape), fully artifact-derived: unicorn executes the image's real reset path up to the call scaffold (the harness fabricates nothing), then runs the stateful export sequence against one wasmtime instance. Symbols from the ELF symtab, never disasm text.

• origin/main: FAIL (4 divergences) — get_lo→0x20020000, get_hi→0x81 (vector-table garbage through R9=0), i32 canary wrong
• this branch: PASS — startup R9=0x20010000; init reads 0x9abcdef0/0x12345678/0xc0ffee1 before any set; i32-declared-after-i64 offset canary at the summed-width slot; fix(globals): #643 thumb-2 i64 global.set/get — register-pair lowering + type-aware slot layout #645 set-then-get pin still green

Gates

• Frozen anchors 10/10; global-free module ELF byte-identical to origin/main; --native-pointer-abi i32 output byte-identical, wide-global refusal still loud
• thumb-2: i64 global.set/global.get truncate to 32 bits (high word silently dropped) #643 differential oracle: PASS (both ARM paths + RV32 loud-skip contract)
• cargo test --workspace green (new: decoder i64-init capture test, startup-materializer byte test, table-layout test)
• cargo fmt --check + cargo clippy --workspace --all-targets -- -D warnings clean

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 96.17486% with 7 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
crates/synth-cli/src/main.rs	95.42%	7 Missing ⚠️

📢 Thoughts on this report? Let us know!