onion-mc/analysis/client-perf

Client performance analysis — Phase 1 (static baseline)

Goal: map the client-side optimisation surface of Onion MC (212 mods, MC 1.21.1 / NeoForge 21.1.251) and rank the levers before benchmarking. Nothing here has been measured on a client yet; Phase 2 benchmarks it (see ../../phase0/ for the fleet harness).

Start with LEVERS.md: the prioritised lever list, plus add and remove candidates.

Layout

Path	What
LEVERS.md	Synthesis: ranked levers, per-machine expectations, add/remove candidates, corrections
proposed-configs/	Phase 1b: staged tuned configs (tiers + bench variants, generated by build.py) and the delivery-mechanism decision (README §1)
BENCH-MATRIX.md	Phase 1b: the Phase 2 run sheet (world, scenes S1–S6, config IDs, blocks B0–B13, decision rules)
bench/onion-bench/	Datapack for the bench world: scene teleports, /fill setups, sound storm, 120 s tour
SPAWN-TUNING.md	Shipped server spawn tuning (Alex's Mobs, Fowl Play, Critters): evidence and the B14 measurement plan
PLAYER-TUNING.md	Phase 1b: player guide (tiers, JVM/heap, shaders, resource packs, Deck)
data/jarscan-report.md	Static scan of every pinned jar: per-frame mixins, per-frame event hooks, atlas/model/shader/sound payloads, entity counts
data/jarscan.json	Raw scan output (re-generate, don't hand-edit)
data/observed-default-configs/	Generated defaults from the last real client launch (2026-09-24, older 110-mod pack). Use these to see actual default values
research/r1-render-core.md	Sodium / Iris + shaderpacks / Flywheel / ModernFix / ImmediatelyFast / EntityCulling / mipmaps
research/r2-dh-xaero.md	Distant Horizons 3.3.2 client + Xaero's minimap / world map
research/r3-entity-animation.md	EMF/ETF + Fresh Animations, GeckoLib, player animation mods, LambDynamicLights
research/r4-effects-audio-hud.md	Sound Physics, AmbientSounds, Subtle Effects, Perception, WATUT, TF / Alex's Caves visuals, HUD mods
research/r5-candidate-mods.md	Perf mods not in the pack (availability on NeoForge 1.21.1, compat, verdict)
research/r6-jvm-platform.md	GC / heap / JVM, Mesa / NVIDIA driver, Steam Deck platform levers
research/r7-sable-create-client.md	Sable / Aeronautics / SSRD / Create + Flywheel client cost
research/r8-sweep-content-mobs.md	Known-issue sweep: mob and content mods
research/r9-sweep-libs-ui-misc.md	Known-issue sweep: libraries, UI, misc client hooks
scripts/jarscan.py, scripts/report.py	The scanner
scripts/jx.py	Jar inspection without unzip/java: ls, cat, str (class-file string constants); nested jars as outer.jar!inner.jar

Re-run the scan after the mod set changes:

# jars: download every mods/*.pw.toml url into a dir (analysis/audit/fetch.py does this)
python3 scripts/jarscan.py /tmp/onion/jars ../.. data/jarscan.json
python3 scripts/report.py data/jarscan.json data/jarscan-report.md

Method and confidence

• Static scan. The scanner parses class-file annotations to get real @Mixin targets and injector kinds. It reads constant pools for event-type references and PNG headers for texture sizes, and it skips the Fabric-only payloads of multi-loader jars. The numbers are surface, not cost: a mixin on LevelRenderer may be a one-line early-return.
• Research. Nine subagent briefs (deepseek-v4.1-flash / glm-5.3-flash) are tagged VERIFIED / DOCUMENTED / COMMUNITY / INFERENCE with URLs. Load-bearing claims were spot-checked by the orchestrator. Corrections are listed in LEVERS.md §7.
• Observed defaults. These come from the old Prism instance's generated configs. They are the authoritative "what a player gets today" values for the mods that existed then.