feat(passes): leverage snarkVM aggregate ternary opcodes

[mohammadfawaz]

Summary

Switch the snarkvm dependency to the mohammadfawaz/complex_ternary branch (snarkVM PR #3222), which extends the ternary instruction to accept literal, array, and plaintext struct operands in addition to registers. The Leo flattening pass is updated to take advantage of this:

• Aggregate ternaries over arrays, plaintext structs, primitives, identifiers, and literals are now preserved in emitted bytecode rather than decomposed field-by-field.
• Records are still decomposed per-field, because their ciphertext components cannot appear directly as ternary operands.
• The pass docstring is updated to reflect the new behaviour in both transition and finalize contexts.

rand / rand_chacha are bumped to 0.10 to match snarkVM, with the corresponding CLI call-site adjustments.

New compiler and execution tests cover ternaries over structs, nested structs, arrays, arrays of structs, records, primitives, and the identifier type, in both transition and finalize contexts.

Bytecode impact (compiler test suite)

Measured across the 1,408 expectation files under tests/expectations/compiler/**/*.out that are present on both master and this branch:

	Before	After	Δ
Total bytes	1,445,043	1,332,058	−112,985 (−7.8%)
Bytecode lines	—	—	−1,648
Files changed	—	—	20 of 1,408

Top savers (bytes saved):

• option/unwrap_or_deep.out — 64,668
• storage/external_storage.out — 15,020
• storage/aggregates.out — 9,529
• function/flatten_test.out — 6,161
• option/in_modules.out — 3,154
• storage/{primitives,signed,unsigned,external_storage_multi_deps}.out — 2,186–2,492 each
• function/flatten_arrays.out — 1,254
• examples/tictactoe.out — 1,076
• option/implicit_wrapping.out — 1,021

Savings concentrate on code that previously forced field-by-field destructuring of ternaries — struct/array conditionals, option-like wrappers, and external storage aggregates.

Bytecode impact (leo-examples)

Measured by running leo build on each of the 34 programs under examples/ with the master CLI and with this branch's CLI, comparing build/main.aleo byte sizes:

	Before	After	Δ
Total bytes	132,513	131,389	−1,124 (−0.85%)
Programs changed	—	—	1 of 34

Only tictactoe/main.aleo changed: 6,525 → 5,401 B (−17.2%), 207 → 183 lines. The per-row/per-cell decomposition of the Board struct now folds into three aggregate ternary ops over Row/Board.

All other 33 example programs produce byte-identical bytecode — they don't conditionally select over composite types, so the new opcodes have nothing to fold.

Bytecode impact (compliant-stablecoin)

Measured by running leo build on the 5 programs of ProvableHQ/compliant-stablecoin (ported to Leo 4.x) with the base CLI and with this branch's CLI:

Program	Before	After	Δ
merkle_tree	10,394	7,984	−2,410 (−23.19%)
multisig_core	39,624	39,624	0
freezelist_program	6,432	6,432	0
stablecoin_program	18,309	18,309	0
bridge_program	18,601	18,601	0
Total	93,360	90,950	−2,410 (−2.58%)

All savings come from merkle_tree, which is the only module conditionally selecting over composite types ([field; 3] arrays inside the sibling-hash loop). The other four programs compile to byte-identical bytecode — they don't use aggregate ternaries, so the new opcodes have nothing to fold.

Test plan

• cargo test -p leo-passes
• cargo test -p leo-compiler --lib test_compiler
• cargo test -p leo-compiler --lib test_execute (new ternary_* cases)
• cargo clippy -- -D warnings
• cargo +nightly fmt --check