PR 30799 real-environment verification, 2026-11-04, Linux x86_64
main: 9e85538535c8553901351112ab02dd1dd62db5ee
PR: d1d4fa5dcbd5695541788408fe90e6de2f2054f7

Setup:
git worktree add ../main github/main
vp i (main and PR)
setsid nohup vp run dev --home-dir /home/user/verify-disk/t3home > /tmp/main-dev.log 2>&1 &
Stopped main by captured process group before starting PR with the same command.
Main ports: web 8693, server 16733. PR ports: web 5733, server 13773.
Playwright 1.60.0 from repository dependencies, /usr/bin/google-chrome.
Task-owned ext4 loopback mount; real app userdata on that mount.
Created project /home/user/verification-project and thread through web UI.
Provider session persisted model: codex / gpt-6.1-sol, cwd /home/user/verification-project.

Trigger:
Submitted real Codex turns through composer, running sleep 30 then cat README.md.
fsfreeze -f /home/user/verify-disk
40 concurrent dd writes on that mount, each timeout 100, bs=4096 count=1 conv=fsync.
External root-filesystem script thawed after 70 s, with EXIT trap to thaw on failure.
Observed server main thread D-state, percpu_rwsem_wait, during disk freeze.
No telemetry injection, mocks, fixture database, or replay provider.

Client-received lastEventLoopStall:
main: {durationMs:69934,loadAverage:30.5,cpuCount:8}
UI: Its CPU is overloaded (load 31 on 8 cores).
PR: {durationMs:57519,loadAverage:32.4,cpuBusy:0.35,cpuCount:8}
UI: Its processes are waiting on disk or other I/O (load 32 on 8 cores, CPU only 35% busy).

Neighbor: 8 bounded low-priority CPU workers during another disk-stalled real turn.
PR: {durationMs:65371,loadAverage:44.1,cpuBusy:0.99,cpuCount:8}
UI: Its CPU is overloaded (load 44 on 8 cores, 99% busy).
Stopped CPU workers by captured process group after capture.

Recovery:
Disk-wait and CPU-branch turns completed through the client after thaw.
Called existing client hostResources atom through its authenticated RPC:
{cpuUtilization:0.006329113924050667,cpuCount:8,availableMemoryBytes:10702680064,totalMemoryBytes:16768798720}
Another real Codex turn created verification-result.txt, read it back, replied recovery verified.
Filesystem: recovery ok newline. Web diff: +1 line, recovery ok.
Reloaded web client: persisted reply and correct diff still available.

Focused checks on task-owned Blacksmith worker:
vp i
vp test run apps/server/src/observability/EventLoopMonitor.test.ts apps/server/src/resourceTelemetry/HostResources.test.ts packages/client-runtime/src/connection/hostFreeze.test.ts
3 files passed, 9 tests passed. Worker stopped.

Limits:
First PR attempt immediately after submit produced no new stall record. Repeating during a running tool produced the measured I/O notice.
Restricted-affinity attempt likewise produced no fresh stall record; neutral fallback not verified end to end. Affinity restored afterward.
Monitor has an existing sample/reset boundary limitation; these runs do not prove every freeze is reported.
No native mobile, Electron shell, macOS, Windows, cgroup quota, remote relay, or tunnel verification.
Controlled filesystem blocking proves this path, not naturally saturated storage or universal cause attribution.
No source changes, commits, pushes, label changes, or merge operations.