Benchmark — structural-only suites are unchanged in v0.3

v0.3 adds judging, flakiness, and comparison. The binding promise of the epic was that none of it touches the default path: a suite with no llm_judge and no repeat must run at v0.2 speed and cost. This is the load-bearing property that keeps dryfire a deterministic, free, never-flaky merge gate.

Why there is no regression (architectural)

A structural-only suite takes exactly the v0.2 code path:

• No judge stage. Composition wires the judge enrichment callback only when a case actually uses llm_judge. A structural-only run passes judge=None, so the scheduler's judge branch is never entered and no judge cost is accounted.
• No repetition expansion. With repeat: 1 (the default), the scheduler produces one unit per case and the aggregation returns the single result unchanged — byte-identical to v0.2. No pass-rate machinery runs.
• No judge cost channel. judge_usage/judge_cost stay at their zero/None defaults and never appear in the output; a structural-only trace serialises identically to v0.2.

There is, quite literally, no new code in the path.

Empirical (reproducible)

$ uv run python scripts/benchmark_structural.py 50
cases          50
passed         50
exit code      0
wall clock     ~240 ms  (~4.8 ms/case)
judge activity 0 cases  (must be 0 — structural-only fast path)
total cost     0.0  (offline → 0)

The script (scripts/benchmark_structural.py) runs a 50-case provider: fake structural-only suite and asserts the fast-path guarantees: exit 0, zero cases touched the judge channel, and zero cost. The wall-clock is dominated by process/loop overhead, not by anything v0.3 added — judging and repetition contribute exactly nothing when unused.

(Numbers are machine-dependent; the invariants — judge activity 0, total cost 0 — are not.)