atlas: disable keep-alive on the test/proxy LLM client to fix Premature close

[jpr5]

What

Disable HTTP keep-alive on OpenAIDistiller's client only when a baseURL is configured (in this repo, always a local aimock/proxy server in tests). Production (no baseURL → real OpenAI over HTTPS) keeps the SDK's default keep-alive agent. One-file change in src/atlas/llm.ts.

Root cause

The Static Quality test job went green on 2026-06-23 and red on 2026-06-25 with no material code change — it is a load-dependent flake that turned deterministic under CI's contention, not a date/time bomb.

The Atlas LLM seam (OpenAIDistiller) is exercised against an in-process aimock server through the OpenAI SDK, whose node-fetch transport pools keep-alive sockets. Under the full suite, every aimock-backed test file runs its own server in parallel — doubled by the committed-then-rebuilt dist/ copy. A server can close an idle pooled socket exactly as a request reuses it, which node-fetch surfaces as FetchError: ... Premature close. This affected all aimock-backed suites (atlas-llm, atlas-exclude, atlas-adapter-episodic, atlas-artifact-sync), including the 9 atlas-artifact-sync failures in this report.

Pinning a non-keep-alive http/https Agent for the mock/proxy case forces a fresh socket per request, removing the reuse race.

Verbatim RED (from CI run 28190002185, 2026-06-25)

FAIL src/__tests__/atlas-artifact-sync.test.ts > syncApprovalArtifact (S17) > rejects a checked candidate that an english exclusion rule drops (via aimock)
FAIL src/__tests__/atlas-artifact-sync.test.ts > syncApprovalArtifact (S17) > feeds a checked candidate's CHILD-BLOCK prose to english rules ...
FAIL src/__tests__/atlas-artifact-sync.test.ts > syncApprovalArtifact (S17) > stamps a DATE-ONLY freshness.as_of on a reconstructed badge-less row (X24)
... (9 failed in src + 9 in dist copy)
FetchError: Invalid response body while trying to fetch http://127.0.0.1:45235/v1/chat/completions: Premature close
AssertionError: expected [Function] to throw error including '[atlas/llm] expected a JSON object fr…' but got 'Invalid response body while trying to…'
  Received: "Invalid response body while trying to fetch http://127.0.0.1:37779/v1/chat/completions: Premature close"

Local RED-GREEN proof (real surface: OpenAIDistiller + real aimock server + OpenAI SDK transport)

The race is environment-sensitive (deterministic on CI's contended Ubuntu runner, invisible on a fast 18-core dev box under the natural suite). To prove the fix on the real failure surface, a harness drove a real OpenAIDistiller against a real LLMock through a TCP proxy that reaps idle keep-alive sockets — modeling a server closing a pooled socket as the client reuses it. The ONLY variable is the keep-alive setting the fix toggles:

RED  (keep-alive ON, pre-fix transport):  mode=keepalive ok=23 failures=17   ← socket-reuse race fires
GREEN (non-keep-alive Agent, the fix):    mode=fixed     ok=40 failures=0     ← race eliminated

The SDK masks the torn-socket-on-reuse as "Connection error" locally; CI's node-fetch transport surfaces the same race as "Premature close" — same root cause, same fix.

GREEN: affected suites (src + built dist), post-fix

$ npx vitest run "atlas-artifact-sync" "atlas-llm" "atlas-exclude" "atlas-adapter-episodic"  (src + dist)
 Test Files  8 passed (8)
      Tests  156 passed (156)

The 9 atlas-artifact-sync failures (and the parallel dist copies) all pass.

Other Static Quality gates (run locally, all green)

prettier --check          → All matched files use Prettier code style!
tsc --noEmit -p tsconfig.scripts.json → (no errors)
npm run build             → exit 0
check-version-sync.sh     → ✓ Versions in sync: 1.15.4

Scope

One file (src/atlas/llm.ts). No changes to notification workflows, RAG config, or anything unrelated. dist/ is gitignored and regenerated by CI's npm run build, so no committed compiled copy to sync.