cli: --bench --json envelope with top-level engine field

[danieljohnmorris]

Summary

Slim slice of adoption brief 4. --bench previously ignored output mode and dumped text only. With --json it now emits one JSON object per engine measurement, top-level engine field naming which engine produced the numbers — so agents can filter bench results without parsing prose.

$ ilo 'add a:n b:n>n;+a b' --bench add 1 2 --json
{"schemaVersion":1,"engine":"tree","result":"3","iterations":10000,"totalMs":6.83,"perCallNs":683}
{"schemaVersion":1,"engine":"vm","variant":"fresh","result":"3","iterations":10000,"totalMs":1.55,"perCallNs":155}
{"schemaVersion":1,"engine":"vm","variant":"reusable","result":"3","iterations":10000,"totalMs":0.66,"perCallNs":66}
{"schemaVersion":1,"engine":"jit","result":"3","iterations":10000,"totalMs":0.43,"perCallNs":43}

The other deliverables in the brief either already exist (engine docs at skills/ilo/ilo-engines.md) or were deliberately reversed (the --engine CLI flag, removed in 0e30891). This PR is just the JSON envelope.

What's in the diff

• cli: --bench honors --json with per-engine envelope — run_bench takes a json flag from the resolved OutputMode. Each engine block emits either text or one JSON line. Python comparator + Summary skipped in JSON mode (Python isn't one of the four engines; Summary is prose).
• tests: cover bench --json engine field; pin eval_inline bench to --text — new regression_bench_engine_field.rs (4 tests) covers the JSON contract; the 7 pre-existing bench_* tests in eval_inline.rs need --text because subprocess mode now auto-resolves to JSON (matching every other subcommand).
• docs: ilo-engines.md notes the bench --json schema — replaces the misleading per-engine bench table (every --bench invocation already exercises all engines) with the actual invocations and the JSON envelope shape.

Notes

• aot is reserved in the JSON contract but --bench doesn't measure AOT today (use ilo build then time the produced binary). When a future change adds an AOT bench it gets "engine":"aot" for free.
• VM emits two records distinguished by variant: "fresh"|"reusable". Keeping both is useful — the reusable VmState measurement is closer to the steady-state cost agents care about, while the fresh-compile path matches what ilo file.ilo does today.

Test plan

• cargo test --release --features cranelift (all green)
• cargo fmt --check
• cargo clippy --release --features cranelift --tests -- -D warnings
• New regression test covers tree / vm (fresh + reusable) / jit
• --text mode unchanged

Follow-ups

• Add a bench measurement for the AOT engine so "engine":"aot" is actually produced (separate brief — needs design on whether bench builds + spawns a binary, or compiles in-process).
• Engine-can't-handle-program diagnostic, mentioned in the brief — separate PR, needs the design decision on what to do under the auto-fallback regime.

[codecov]

Codecov Report

❌ Patch coverage is 98.48485% with 1 line in your changes missing coverage. Please review.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/main.rs	98.48%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!