Files
aeternalis/docs/VERIFICATION.md

79 lines
4.1 KiB
Markdown

# Verification performed
## September 9 client regression checks
The initial timeout-only CI correction did not address render-dependent input or
the premature "Connected" status. The existing boss test now holds animation-frame
callbacks after targeting: it reproduced unchanged boss HP before the input-timer
fix and passes afterward against the real authoritative server. The existing
desktop reconnect test holds incoming server snapshots, checks the synchronizing
state, then releases them and requires the original character and all nine chunks.
Neither test substitutes client-calculated gameplay results.
Required TypeScript/build and Go checks pass locally. CI on its Linux software
renderer remains the deployment verification gate; local passing tests do not
establish its result. Routine new Playwright coverage remains deferred; these
changes extend the existing failing regressions.
## Original verification baseline
Local checks on Windows, 2026-09-05. Node 24.13.0, Go 1.27.1, Blender 4.3.2,
Helm 3.18.6, and headless Microsoft Edge driven by Playwright.
## Passing checks
- Generated Go/TypeScript contract drift check, TypeScript strict checking, and
production Vite build. Main JavaScript: 600.22 KB, 156.65 KB gzip estimate.
- Go tests: deterministic terrain and seams, movement bounds/replay rejection,
chunk interest, gathering/crafting, permanent death across restart, daily cap,
real wall-time accounting under lag, keystone safety, boss victory/collapse,
offline survivors, transfer progress preservation, resource replication.
- Storage tests: exclusive ownership, stale-owner rejection, durable passages,
cross-account denial, duplicate redemption rejection, simultaneous redemption
at two destinations, later-world passages, session revocation.
- HTTP/WebSocket tests: origin rejection, HttpOnly sessions, unknown JSON fields,
initial chunk delivery, private identity exclusion, malformed socket input.
- `go vet ./...`.
- Four Playwright tests: desktop rendering/movement/gather/reconnect; mobile
portrait and landscape controls/layout; HTTP security behavior; boss rendering
and actual server-authoritative boss damage. Canvas pixel variation, loaded
models, triangle counts, and progressing render frames were checked.
- Desktop, mobile, landscape, and boss screenshots visually inspected.
- Helm lint and rendering: five distinct Deployments/Services/Ingresses plus a
NetworkPolicy. No cluster resources were applied.
## Initial concurrency measurement
`go run ./cmd/load -url http://127.0.0.1:18090 -clients 20 -duration 15s`
Twenty authenticated WebSocket clients on an isolated local development world,
with periodic movement/idle intents and snapshot reads. Join setup is excluded
from the throughput window. This is a short connection/replication smoke test,
not a long-duration soak, active combat workload, mobile hardware benchmark, or
large final-boss crowd test.
| Measurement | Before static-resource optimization | After |
|---|---:|---:|
| Concurrent clients | 20 | 20 |
| Measurement duration | 15 seconds | 15 seconds |
| Delivered snapshots | 1,500 | 1,500 |
| Outgoing application payload | 7,548.5 KiB/s | 560.3 KiB/s |
| p95 interval between snapshots | 206.9 ms | 203.0 ms |
| Unexpected disconnections | 0 | 0 |
The target snapshot interval is 200 ms. Sending static resource transforms once
per chunk reduced measured outgoing payload by about 92.6%. Figures exclude TCP,
WebSocket, and TLS framing. CPU/RAM/GPU saturation and 2,000 players remain unproven.
## Not executed locally
- Docker image build/runtime: Docker Desktop's Linux daemon was not running.
- PostgreSQL-backed tests: no test database was configured locally; the workflow
provides a PostgreSQL service and runs the same storage lifecycle suite there.
- Go race detector: no local C toolchain was available; the Linux CI job runs it.
- Real OIDC-provider login, cross-host production transfer, TLS ingress, network
policy enforcement, image publication, ArgoCD sync, and Kubernetes rollout.
The Playwright Chromium download timed out; all browser tests ran successfully
using installed Edge instead. CI is configured to install Chromium.