Performance Evidence and Measurement

What improved across 7.x

Factory reduced repeated hydration, parameter copies, graph scans, venue/participant assembly and tally work. These are historical paired measurements, not fresh benchmarks of the current branch. Release membership follows the first 7.x tag containing the change. See the 7.x overview for the short account.

First release / status	Change and workload	Recorded before → after
7.5.0	#5084, #5086: fixed 2,026-call scoring workload	Hydrations 6,034 → 3,408 (43.5% fewer); median 2 → 1, worst 14 → 8. positionTargets calls 89,071 → 20,459 (77.0% fewer).
7.5.0	#5098: 456 in-context lookups / TEAM-line scores	Lookup total 926 → 142 ms (6.5×); line median 12.95 → 8.88 ms (31.4% lower). Downstream target calls 12,059 → 9,874 overlap this workload.
7.10.0	#5356: 168 scheduled / 504 total matchUps	Readiness about 1,270 → 5 ms; participant rest about 1,290 → 26 ms after avoiding unused parameter copies. Workload timings, not every individual call.
7.15.0	#5469: raw ID location before hydration	Lookup hydrations: FMLC128 223 → 1, TEAM16 64 → 4, multi-event 226 → 1. Format lookups: getMatchUpFormat 222 → 0 (FMLC128), getTieFormat 64 → 4 (TEAM16) and 229 → 4 (multi-event). Action hydrations 444 → 222 / 120 → 60 / 279 → 54.
7.15.0	#5471: bare engine import	27.1 → 15.7 KB gzip (42.1% smaller); bundled modules 124 → 69. Not the full application bundle.
7.15.0	#5482: FIC128 invocation indexes	addGoesTo 11.6 → 1.7 ms (6.8×); dependencies 21 → 1 ms (21×).
7.16.0	#5503: venue reuse	Earlier profile built venue data 511 times / performed 8,176 court operations; now shared per invocation. Roughly 2× prototype timing is not a final production result.
7.16.0	#5505: format/tally memoization	FMLC128 432 → 38.5 ms (11.2×); COMPASS64 176 → 23 ms (7.7×); SE32 46 → 4.9 ms (9.4×).
7.16.0	#5506: shared draw context	COMPASS context built once instead of eight times. Prototype 333 → 69 ms is not an isolated final-release timing.
7.16.0	#5507: DOUBLES64 with 2,048 unrelated pairs	Lookup 485 → 56 ms (8.7×); complete measured read 548 → 168 ms (3.3×).
7.16.0	#5522: SE512 tally and then-current read	Tally 546 → 6.9 ms (79×), calls 1,033 → 58; public read 413 → 72 ms (5.7×). Other reads: FMLC 45 → 21, COMPASS 24 → 14, SE64 13 → 6 ms.
7.22.0	#5779, #5788, #5796, #5798: setMatchUpStatus settle passes ask the structural index before hydrating; census hydration gate (verify:hydrations)	Draw hydrations a call: BYE matrix, DOUBLE_WALKOVER 8.7 → 3.0; WALKOVER propagated 9.1 → 3.4; BYE correction routes 7.6 → 1.9. Work counts from the gate's baseline at each tag, not latency.
7.23.0	#5804, #5805, #5807, #5809, #5834: the remaining setMatchUpStatus settle passes ask the structural index before hydrating	Draw hydrations a call against 7.22.0: BYE matrix, DOUBLE_WALKOVER 3.0 → 1.6 (−47%); WALKOVER propagated 3.4 → 1.5 (−55%); exit matrix, DOUBLE_WALKOVER 4.4 → 1.9 (−56%); deep BYE 1.9 → 1.3 played out, 3.7 → 2.5 corrected (−33%). Work counts from the gate's baseline at each tag, not latency.
7.23.0	#5820, #5821: direct scheduling's per-placement costs; a scheduling write's matchUp lookup reads the stored matchUp	COMPASS 128 direct scheduleMatchUps (312 matchUps, 16 courts, one policy): 81% less time than 7.22.0, placements identical. Paired run on one idle host, versions alternated (Button, 2026-10-26).

Elimination projection changed after the tally benchmark

#5525 removed elimination tallying from public projections. Thus #5522's 72 ms public read is historical, not a current-head benchmark. Its standalone tally optimization remains useful, but does not explain a projection that skips tallying. RR refresh also became group-scoped and policy-aware. See the results contract. Do not multiply gains or infer a new time without measuring the final candidate.

Choosing the work a consumer needs

• Use a matchUp ID filter for a targeted read; full structural facts still support completion, progression and TEAM-parent context.
• Use hydrateParticipants: false when entry metadata suffices. Thin output retains scoring/context facts; it is not raw-state access.
• Omit recovery calculations when unused. contextProfile.exclude redacts final output; it does not promise that all computation associated with the excluded attributes is skipped.
• Use structuresProfile: 'STUBS' where supported for a smaller response. Structure/group dependencies can still require full computation.
• Prefer named governor imports over the aggregate governors object. See bundle measurements.

These caches/indexes belong to one invocation/traversal, not a durable cache across mutations. Do not retain them across writes without owning invalidation. Narrow rollback snapshots require a proven mutation footprint and complete restore evidence before optimization.

Reproducing and extending the evidence

Build with pnpm build. pnpm verify:consumer-bundle checks bare-engine, named-query, aggregate-query and full-engine fixtures using esbuild bundle/minify/tree shaking, ESM output and gzip level 9. scripts/verify/baseline/consumer-bundle.json budgets raw/gzip growth. A budget is not a fresh timing benchmark. Compare releases using identical fixture bodies and imports.

For timing, pin code, Node and dependency revisions; reuse identical fixtures and histories across versions and record their hashes. Warm up, take enough repetitions for median/p95, control competing load and verify result parity. Separate initialization, operation, serialization and consumer render. Record hydrations/parses/traversals, output bytes and allocations with elapsed time. Counts measure work, bytes measure payload and milliseconds measure latency; none substitutes for the others.

A useful panel is 7.0/7.5/7.10/7.15/next candidate over SE32/128/512, FMLC128, COMPASS64, TEAM16, RR64, multi-event and unrelated-participant-heavy fixtures. Include scheduled/unscheduled and one-ID reads, refresh on/off, line outcomes and rollback-enabled queues. Report regressions as well as gains. Historical timings exclude network, storage and browser rendering. Gains overlap and cannot be added or multiplied; no end-to-end speedup for the whole 7.x cycle has been established.