erik/acdream
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

test(app): make canonical soak comparisons workload-aware

Browse source 
Pin every ACE teleloc to an explicit starting quaternion and distinguish stable-workload owner growth from legitimate authoritative population, visibility, and cache-retirement changes. Preserve the original process residency and frame-cost limits while making canonical deltas actionable.

Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
Erik 2026-07-22 20:13:29 +02:00
parent bca4148739
commit 6c5e0604c9
4 changed files with 88 additions and 44 deletions
Download patch file Download diff file Expand all files Collapse all files

23
docs/plans/2026-07-22-gamewindow-slice-8-checkpoint-k-canonical-soak.md
View file

@ -136,17 +136,18 @@ Every checkpoint hard-gates zero:
- staged mesh uploads and bytes;
- composite warmup work.
The same-location Caul return → Caul plateau comparison requires exact equality
for deterministic cache/allocator state: loaded/visible/total landblocks,
teardown/retirement queues, mesh render-data/atlas/estimated bytes, staged work,
tracked GPU bytes/buffers/textures, and composite/particle texture ownership.
Connected `WorldEntities`, `AnimatedEntities`, `LiveEntities`, and materialized
live entities may legitimately change with authoritative population; those
deltas are named workload warnings. VFX binding/owner/script/light counts remain
hard equality gates when that population is stable, and become named workload
warnings when the authoritative population changed. Particles, emitters, and
active scripts are always transient workload warnings rather than silent
exclusions.
Every route teleport supplies an explicit identity quaternion so the timed turn
begins from a reproducible heading. The same-location Caul return → Caul
plateau comparison requires exact equality for the deterministic loaded/total
world domain, while visible-landblock and authoritative entity/animation/live-
entity changes are named workload warnings. Teardown/retirement queues, staged
work, and composite warmup remain hard zero at every checkpoint. Mesh, atlas,
tracked-GPU, composite/particle-texture, and VFX owner counts may retire and
shrink normally; any growth is a hard leak failure when the visible and
authoritative workload is stable, and a named owner-growth warning when that
workload changed. Particles, emitters, and active scripts are always transient
workload warnings rather than silent exclusions. The report therefore says
which canonical owner grew without mistaking legitimate retirement for a leak.
Managed used/committed deltas are reported diagnostically. Existing
working/private-memory, update-p95, and allocation-p50 thresholds remain byte-
for-byte unchanged as the secondary residency/performance guard.