Phase 13. LLM generate with cassette replay

Field	Value
MEP	MEP-53 §Phases
Status	LANDED
Started	2026-05-29 (GMT+7)
Landed	2026-05-29 (GMT+7)
Tracking issue	#22604
Tracking PR	#22499
Commit	819cb38daa

Gate

TestPhase13LLM walks tests/transpiler3/rust/fixtures/phase13-llm/ (11 fixture subdirectories) and asserts byte-equal stdout. Each fixture is its own subdirectory with a cassette/ folder containing {sha256_hex(provider:prompt)}.txt files; the test sets MOCHI_LLM_CASSETTE_DIR to that path before running the binary. Coverage: generate openai, generate anthropic, generate llama, model-override variant, multi-step prompts, generate-in-fun, generate-in-var, prompt concatenation.

Lowering decisions

let r = generate openai {
    prompt: "hello"
    model: "gpt-4o"
}

lowers to:

let r: String = mochi_runtime::llm::call("openai", "hello");

The model field is currently ignored in the runtime (it would be sent over the wire to the provider's API, which is itself out of scope for the cassette-replay-only mode). The provider name is folded into the cassette key alongside the prompt.

mochi_runtime::llm::call(provider, prompt):

1. Reads MOCHI_LLM_CASSETTE_DIR from env. Raises panic code 99 if unset.
2. Computes key = format!("{}:{}", provider, prompt).
3. Hashes key with the inline SHA-256 implementation (avoids the sha2 crate dep).
4. Reads {dir}/{hex(hash)}.txt. Raises panic code 99 if missing.
5. Trims trailing whitespace (matching vm3 behavior) and returns.

Live calls to actual provider APIs are out of scope for MEP-53. The cassette-replay design lets the fixtures be fully reproducible without network access; a future sub-phase could add a live mode gated by MOCHI_LLM_LIVE=1 for development.

Files changed

File	Purpose
transpiler3/rust/lower/llm.go	generate expression lowering
runtime3/rust/mochi-runtime/src/lib.rs	Add llm module + inline SHA-256
transpiler3/rust/build/phase13_test.go	11-fixture-subdir gate, sets MOCHI_LLM_CASSETTE_DIR
tests/transpiler3/rust/fixtures/phase13-llm/*/{*.mochi,*.out,cassette/*.txt}	11 fixtures

Test set

• TestPhase13LLM/<fixture> for each subdirectory (11 fixtures).

Closeout notes

The choice to inline SHA-256 (~80 LOC) rather than pull in the sha2 crate has two motivations: (1) it keeps the runtime crate dep-free past itertools for fast cargo builds, (2) it lets the embedded feature gate work cleanly — sha2 brings in cpufeatures which doesn't compile on no_std without alloc + portable-atomic. The inline implementation is the classic FIPS 180-4 algorithm with K constants and round function inlined; performance is irrelevant because the hash runs once per LLM call.

The trim-trailing-whitespace step matches vm3's behavior: vm3 strips trailing \n from cassette files when reading. Without this, a Mochi print(generate ...) would emit the LLM response plus an extra blank line.