bench-baselines

Committed performance reference points for the phase gate described in BENCHMARKING.md ("Baselines: capture & compare") and docs/audit/2026-06-architecture-audit.md (Phase 0.0).

Conventions:

• phase-<N>-pre.json — captured on the commit a refactor phase branches from, committed with the phase's first PR. Every PR in the phase runs bench-baseline compare against it.
• phase-<N>-post.json — captured at phase close on the same hardware class and env knobs (FLUREE_BENCH_PROFILE, FLUREE_BENCH_SCALE) as the pre-baseline. All tuples must be within budget vs pre; improvements are banked by tightening regression-budget.json in the closing PR.
• Ad-hoc local baselines (validation runs, experiments) should live outside the repo or be cleaned up before merge — only phase reference points belong here.
• Each baseline records a host block (class, name, os, arch, cpu model, cores). host.class names a comparability set, from FLUREE_BENCH_HOST_CLASS or the derived {os}-{arch} default. compare refuses (exit 2) when the baseline's class differs from the comparing host's, since neither wall-clock time nor a tracking allocator's peak survives a machine change; --allow-host-mismatch downgrades those to non-gating annotations. Per-phase share drift gates either way — it is a ratio within one run, so the machine cancels.
• A baseline that also carries a corpus block (sha256, byte length, element count, input/output syntax, thread count) makes compare refuse outright on a corpus change. There is no override: comparing different inputs is never meaningful, and the result would look exactly like a real regression.
• Baselines built with capture --accumulate over ≥ 5 runs carry a median + MAD per scenario, and the gate widens each budget to $3 \times \text{MAD} / \text{median}$ when that exceeds the declared budget. Prefer this for any baseline meant to gate on shared CI runners.

guardrails-pre.json is the PR-1 (guardrails net) reference — captured at the merge-base engine, quick profile, tiny+small, schema 1. It predates the host block, so its runner_class: "local" is read as host.class = "local": an honest "some developer machine, unspecified". That is why the per-PR compare still runs with --allow-host-mismatch.

Capture (the --label/--out name the phase reference; capture at the merge-base commit):

cargo run -p fluree-bench-support --bin bench-baseline -- \
    capture --label phase-1-pre --out bench-baselines/phase-1-pre.json