Make hanging-request fake-server teardown deterministic and bounded

[HelloThisWorld]

Summary

A post-merge main run on macOS/node 20 spent the full 30-second Vitest budget inside ollama.test.ts > a timeout aborts the request deterministically — a test whose own contract is ~1.5 seconds — while 1,649 of 1,650 tests passed (failing job). The test drives the fake server''s timeout behavior (accept the request, never respond) and closes the server in a finally, so the anonymous outer timeout could not even distinguish a hung client abort from a hung fixture teardown.

Inspection found the concrete design gap on the teardown side: close() relied on closeAllConnections?.() plus the server close callback, with no bound on the wait and no tracking of the socket a client abort leaves mid-unwind — a code path that could wait forever on an OS/runtime scheduling hiccup. Both fake HTTP servers (Ollama and OpenAI-compatible, which shared the identical pattern) now use one hardened lifecycle, trackedServerLifecycle: every accepted TCP socket is tracked from the connection event; close() stops the listener, explicitly destroys every live socket (idle keep-alive and intentionally hanging requests), and bounds the whole wait — on expiry it force-destroys stragglers and rejects with <server> teardown timed out after N ms with K tracked socket(s) still open, so any future hang fails fast at the actual location with actionable diagnostics instead of hiding behind the outer test budget.

Deliberately unchanged: the timeout fixture still leaves a real hanging HTTP request that only the client''s AbortSignal.timeout ends (the abort path stays genuinely tested); production timeout/cancellation semantics in safeHttpRequest and the Ollama runner are untouched; and no Vitest timeout budget was raised anywhere — the point is to remove the hang, not to hide it. Loopback-only traffic, no platform-specific behavior.

New regression tests (fake server lifecycle describe block): teardown after a client-aborted hanging request completes well inside a 2s bound with the timeout still classified timed-out/timed_out; teardown while a request is still in flight destroys it and settles; ten abort→teardown cycles stay bounded. The adjacent cancellation test (same hanging fixture) is covered by the same hardened helper.

Validation

• tests/runners/ollama.test.ts: 27/27 (24 existing + 3 new)
• Stress: 20 consecutive runs of the whole file (~20 timeout + ~200 external-cancellation abort→teardown cycles): 0 failures
• Every consumer of the changed helpers (openai-compatible, mcp-runner-tools, openai-execution, multi-runner, conformance, cli-v06-runner): 90/90
• pnpm lint / pnpm typecheck / pnpm build: pass
• pnpm test: 116 files, 1,651 tests, exit 0
• macOS node 20 targeted validation: NOT RUN (no macOS environment here; validated on Windows — the hardening is platform-neutral by design)

Checklist

• pnpm lint, pnpm typecheck, and pnpm test pass locally
• pnpm check:public-contracts passes — no contract touched (test helpers only)
• Documentation updated for any user-visible behavior change — none
• CHANGELOG entry added for any user-visible change — none
• Everything is in English (code, comments, docs, commit messages)
• No employer or client proprietary content — examples and fixtures are synthetic
• No credentials, tokens, or secret values anywhere in the diff, fixtures, or recorded test output