Lifecycle-aware WSL cache reclaim — Primary Node live qualification

Date: 2026-07-18

Host: Primary Node, Windows + Docker Desktop WSL2

Hardware: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition (95,587 MiB)

Feature base: origin/main 8ad73ff4a8b327e50f5ec04d125214235aa64bed

Decision: ADR-0023

Result

The default-off host.toml policy was enabled on the operator machine and a confirmed, manifest-owned serves up heavy loaded the already-cached pinned GPT-OSS Puzzle 88B deployment. The lifecycle hook attributed 49.9 GiB of page-cache growth to the operation, observed settled growth of 0.00 GiB/s, ran the page-cache-only primitive, and reported:

cache reclaim after serves up: reclaimed
(cache 1.1 GB -> 1.1 GB, operation growth 49.9 GB,
 latest growth 0.00 GB/s; distro docker-desktop)

After reclaim, the direct model endpoint remained healthy, advertised the exact expected model, kept its GPU allocation resident, and completed a small inference. The serve and router were then restored to their exact initial stopped state. No checkpoint was downloaded, no promotion or rollback ran, no route changed, and no Puzzle/Gemma configuration was modified.

Policy and target

The operator file did not exist before the run, so there was no prior file to back up. It was created at C:\Users\operator\.anvil-serving\host.toml and intentionally left enabled:

schema_version = 1

[cache_reclaim]
enabled = true
distro = "docker-desktop"
threshold_gb = 16

The managed target was the Puzzle Heavy service already promoted on main:

Property	Value
Compose service / container	heavy / vllm-gptoss-puzzle-heavy
Endpoint	http://127.0.0.1:30002
Model / served name	nvidia/gpt-oss-puzzle-88B / gpt-oss-puzzle-88b
Checkpoint revision	9c0e0746a0d2218b28cc7b2cb3ce4e1a2f50fdb2
Image	anvil-vllm:gpt-oss-puzzle-485463b3498ed3ffcf0c8fcb52c1670a21be5d82
Recipe	Pinned Puzzle 88B recipe

The dry run resolved the enabled policy, docker-desktop, 16 GiB threshold, and the exact managed Compose action before apply.

Measurements

Linux values are direct /proc/meminfo samples from docker-desktop. GiB conversions use 1,048,576 KiB per GiB. vmmemWSL is the Windows process working set and is reported separately; it is not treated as a page-cache measurement.

Point	Linux Cached	SReclaimable	vmmemWSL working set	PRO 6000 VRAM	Router / Heavy
Before load	1,142,288 KiB (1.09 GiB)	25,892 KiB	3.40 GiB	510 MiB	exited 137 / exited 137
After automatic reclaim, model resident	1,275,948 KiB (1.22 GiB)	525,268 KiB	8.39 GiB	93,839 MiB	exited 137 / running
After state restoration	1,148,540 KiB (1.10 GiB)	498,760 KiB	4.94 GiB	510 MiB	exited 137 / exited 137

The hook's internal samples saw 49.9 GiB of operation growth before the drop and approximately 1.1 GiB after it. The later 1.22 GiB sample includes endpoint verification and inference activity. The result materially reduced page cache while leaving the loaded model's GPU VRAM allocation intact.

Readiness and inference evidence

• GET /health returned HTTP 200.
• GET /v1/models returned HTTP 200 with exactly one id: gpt-oss-puzzle-88b.
• A low-reasoning chat-completions request returned HTTP 200, model gpt-oss-puzzle-88b, finish_reason: stop, and visible content exactly cache reclaim live check passed (76 prompt tokens, 32 completion tokens).
• A deliberately smaller 64-token request first ended with finish_reason: length and no visible content because the reasoning model exhausted that budget. It was not counted as the functional success; the model-aware low-reasoning request above is the retained smoke result.

This validates service continuity and residency across the cache drop. It is not a new quality benchmark and does not change the existing Puzzle promotion evidence.

Caveats

• The pre-existing stopped anvil-router container belonged to the Puzzle worktree's Compose project, while the installed operator-home router action tried to create the same fixed container name from a different Compose project. Docker rejected that create with a name conflict. Existing serves up behavior continued the Heavy start; the router remained stopped during model qualification, and all probes used the direct loopback model endpoint. Later, its pre-existing restart: unless-stopped policy auto-started the previously exited container. The final state audit caught that change and explicitly stopped router through the Puzzle worktree's exact Compose file, restoring both containers to exited 137. This is an environment ownership and restoration caveat, not a cache-reclaim failure.
• Reclaim is VM-wide. A different WSL workload may need to reread evicted clean pages, which is why the product default remains disabled and the operator policy is explicit.
• The measurement establishes the pinned Puzzle managed-load path on this workstation. It does not claim request-triggered, voice, ComfyUI, ad-hoc Compose, or non-Windows applicability.