2.7 KiB
Linux canonical late-slope diagnostic — completed, no cold-cap breach
One fixed Linux ARM64 diagnostic ran from source
b18f97236a033b54945b3b9a1d5e306e11e278a7 with OPC UA SDK 2.0.1,
HistorianGateway Client/Contracts 1.0.0 and the pinned self-contained .NET 10.0.14
runtime. The runtime mapping, descriptor controls (2/2), workload (1/1),
shutdown and removal of the task-owned container were verified. The canonical
phase completed 7,201.014 seconds, 1,937 cycles and 484 restarts. The workload,
container and launcher exited 0. The diagnostic receipt explicitly reports
acceptance=false.
No resource assertion latched. Cycle 3 set the original RSS baseline at 225,640,448 bytes (215.19 MiB); the unchanged +128 MiB limit was 359,858,176 bytes (343.19 MiB). The observed quiescent peak was 355,459,072 bytes (338.99 MiB) at cycle 251 / 929.817 seconds, 4.20 MiB below the limit. Retained managed memory peaked at 20,410,640 bytes against its cycle-3 19,923,288-byte baseline; total descriptors peaked at 195 against a baseline of 191. All 1,937 quiescent samples had zero PKI descriptors and zero live session/subscription/continuation/monitored-item inventories.
| Fixed window (minutes) | Samples | Median RSS (MiB) | Min–max RSS (MiB) | Median retained managed (MiB) |
|---|---|---|---|---|
| 30–40 | 162 | 316.35 | 310.81–321.04 | 19.02 |
| 60–70 | 162 | 317.03 | 309.78–321.66 | 19.10 |
| 110–120 | 160 | 317.20 | 313.43–320.18 | 19.05 |
Ordinary least squares over all 1,452 quiescent observations from 1,800 through 7,200 seconds yields +203,355 RSS bytes/hour (+0.194 MiB/hour) and +52,381 retained managed bytes/hour (+0.050 MiB/hour). The largest gap between those observations was 7.641 seconds. At 1,800 seconds, 484 cycles and 121 restarts had completed, exceeding the earlier restart-step work count without asserting new trace readiness.
This is evidence that sampled RSS largely leveled off after startup in this one tracked Linux run. It does not identify which components own all RSS, bound unsampled transient peaks, prove that growth can never recur, or supersede the earlier Linux cold failures. No EventPipe observer was attached; certificate allocation stack tracking and per-cycle quiescent GC remained enabled. This is not uninstrumented resource acceptance, a matched SDK1/SDK2 comparison, or a resolution of macOS native import retention. The original cold cap remains binding; no warmed-baseline rule was adopted. Issue #799 and G4 remain open.
The selected receipt retains exact identities, raw evidence hashes, exit statuses, window statistics and sample counts. The rehearsal and its earlier failure remain separate.