What

• add --warmup-requests-per-lane as a deterministic alternative to --agentic-cache-warmup-duration
• reject configurations that supply both options
• enforce the count budget independently at each agentic replay lane, including roots, continuations, joins, and DAG children
• keep a global requests_per_lane × active_lanes request cap as a zero-overshoot backstop
• preserve replay-barrier and recycled-root state at the count boundary for the warmup-to-profiling handoff
• retain the existing duration-only path unchanged
• regenerate CLI docs and the YAML config schema

Why

Duration-only cache-pressure warmup makes request depth depend on server speed. The same C1 workload sent 670 warmup requests in one run and 343 in another, advancing the live conversation cursors by different amounts before profiling. A strict per-lane request budget makes the warmup depth stable across runs instead of letting faster lanes consume a shared global budget.

Example: --concurrency 16 --warmup-requests-per-lane 10 sends exactly 10 warmup requests on each lane (160 total). Use --agentic-cache-warmup-duration 600 instead when a time-bounded warmup is desired; the two options are mutually exclusive.

Validation

• focused config, timing, strategy, CLI override, and orchestration suites: 175 passed, 1 skipped
• end-to-end CLI coverage for duration-only and count-only warmup: 2 passed
• uv run pre-commit run --all-files: all hooks passed