[paulbalaji]

Problem

(*[8]byte)(b) passes gosvm check and native tests, then gosvm build fails on both targets: the (*T)(p) rule never looked at the operand, so the backend cast the C slice struct to a pointer. [N]T(s) was rejected by Check outright.

Solution

Both forms now follow Go:

• len(s) < N aborts with the existing bounds trap, no new code
• (*[N]T)(s) is the slice's data pointer, so writes through either view (account data included) see each other; a nil slice gives nil, which only N == 0 reaches
• [N]T(s) copies the first N elements; [0]T(s) reads nothing

Both are plain operations in gc's order: in (*[8]byte)(b)[0] + f(), f runs first, its panic wins, and a b it replaced is the one converted. Converting package initializers run at run time; the compile-time evaluator used to return the slice as the pointer and panic. gosvm features reports slice_array_conversions. ERROR_CONTRACT changelog entry included (announce to Qushta).

Tests

• TestSliceArray*: two programs, 256 inputs each, native Go (host 1.23.2) vs host C, comparing output, writes before a failure and panic lines: len == N, len > N, len < N, aliasing, nil and empty slices, struct/array/generic elements, nested generics, escapes, initializers, evaluation order.
• Real SBF on fogo-v0 and solana-v3: 96 inputs per program, all match, failures roll back.
• Dropping the evaluation-order event or the initializer stop fails the tests.
• The five examples/* gosvm projects, a fresh starter and the 4 reference ELFs build byte-identical to 79cf25c. The frozen corpus's one slice conversion ([24]byte(s)) now compiles, recorded in corpusChanges. make test passes apart from the known TestStdlib* and TestFullSwapFacts host failures.

Linked work

Closes #7. hyperlane-xyz/hyperlane-monorepo#9852 worked around it by copying keys into a [32]byte. Evidence: results/compiler/2026-10-19-slice-array-conversions. Conflicts with the #8 fix on ERROR_CONTRACT and HANDOFF are mechanical.