fix(e2e): prepare attached Ollama for export

[sandl99]

Outcome

The attached Ollama export E2E prepares a ready daemon with an installed model, reuses the host's shared proxy port, and stops the competing installer service before export qualification.

Reason

The fixture treated a two-second curl startup timeout as terminal. Its fresh daemon also had an empty model cache, and it requested a different proxy port from the preceding GPU scenarios. The existing host-wide proxy contract rejects that change.

E2E root cause: attached-daemon export fixture / incomplete startup preparation.
The original failure stopped at readiness. Current validation completes fixture preparation and onboarding, then fails at provider metadata verification. The target is not yet green.

Related issues

Refs #11435. Repairs the export E2E introduced by #11473.

Changes

• Accept curl exit 28 alongside connection refusal within the existing 20-read readiness bound. Move that block into the GPU helper and cover recovery, exhaustion, HTTP failure, and terminated probes.
• Preload qwen3.5:9b directly into the attached daemon on 11439. Inherit the target's PROXY_PORT and verify it in the export.
• Stop the installer service after onboarding so it cannot replace the fixture-owned daemon. Update the owning E2E guide.

The polling limits, per-attempt artifacts, assertion budget, export checks, and stopped-backend refusal remain intact. Mutations are not retried. No product code changes relative to main. General nondefault-port model downloads remain outside this export-fixture repair.

Verification

Current commit: b3ccf627af3b1b253563962b176efcfb1b48dab3, including main e4e6391870bef3834cd30cdd997529e5361f89a5 through a clean signed merge.

• GPU helper and polling support suites — 45 passed after the merge.
• Shared proxy-port contract cases — 3 passed after the merge.
• Normal pre-push publication validation, CLI/plugin builds, and compiler checks — passed using canonical validators. All six branch commits are GitHub verified.
• GPU E2E — 2 passed, 1 failed. OpenClaw and Hermes inference pass. Attached-daemon readiness recovers through exit codes 7, 28, 0; model preparation, onboarding, and service shutdown succeed. First export fails with The live inference provider metadata could not be read or verified. All registered cleanup actions pass; stopped-backend refusal is not reached.
• CI — failed because the unchanged setup-nim-flow-vllm-resume.test.ts test for a pinned provider exceeded five seconds in shard 6. The exact test passes locally in 291 ms; no GitHub rerun or timeout increase was applied. The CLI aggregate and overall check inherit that failure.
• The diff contains no secrets, API keys, or credentials.

Review notes

Main includes #11601, which resolves the preceding SDK sandbox-inventory failure tracked by #11421. The rerun reaches the later provider-metadata stage; its precise failing sub-check is not yet identified. No export assertion is relaxed, and no successful export is claimed.

Advisor run 34680869998 is skipped because CI is red. It produced no current-commit specialist review. CodeRabbit skips draft PRs. There are no submitted reviews or inline threads, and no clean Advisor review or waiver is claimed.

---

Signed-off-by: San Dang sdang@nvidia.com

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

_{Comment @coderabbitai help to get the list of available commands.}

[github-code-quality]

Code Coverage Overview

Languages: TypeScript

TypeScript / code-coverage/plugin

The overall line coverage in commit b3ccf62 in the fix/e2e-attached-oll... branch remains at 96%, unchanged from commit 84a9ea5 in the main branch.

---

_{Updated September 12, 2026 07:17 UTC}