M3 benchmark baseline

Recorded from cargo bench -p corium-query (criterion, default profile) at milestone M3 completion. Dataset: musicbrainz-style graph of 100 artists × 3 albums × 10 tracks (3,400 entities, ~10,400 datoms), built once with indexes and statistics materialized before timing.

Hardware note: numbers are from the CI-class container that produced the milestone; they are a relative regression baseline, not absolute targets. Re-record on hardware changes.

Benchmark	Median	What it measures
point_lookup_unique_attr	1.85 µs	entity by unique attribute (AVET prefix), cached parse
join_heavy_artist_tracks	88.2 µs	4-clause join artist → albums → tracks (VAET reverse-ref prefixes)
aggregate_group_by_country	6.8 ms	whole-dataset 3-hop join with grouping (count + avg per country)
pull_heavy_artist_discography	19.4 µs	nested pull with two reverse-ref levels
as_of_view_range_count	164 µs	AVET value-range predicate count on an as-of view

Observations at baseline:

• Point lookup meets the "sub-ms warm" performance posture from query-engine.md by ~3 orders of magnitude.
• The join and aggregate suites are dominated by per-frame BTreeMap cloning in the executor; batch/slot-based frames are the obvious next optimization if these regress into requirements.
• Adding the two-component VAET prefix (bound attribute + bound ref value) during this milestone took join_heavy_artist_tracks from 2.1 ms to 88 µs and aggregate_group_by_country from ~210 ms to 6.8 ms.