Fix serial review preflight obstruction without inference deadlines #2408

Root cause

In ContextualWisdomLab/fast-mlsirm#2171 run 36237188327 / job 108408520367,
retained artifact 10919666896 shows discovery completed and several probes
finished. The next sequential nvidia_nim meta/llama-3.2-90b-vision-instruct
invocation began at 2026-09-26T19:06:49.173Z without a subsequent outcome
before hosted cancellation at 2026-09-27T01:00:56.653Z. This is startup route
readiness, not gateway inference or Strix scanning. The eight-ready target
means a pending probe prevents reaching otherwise usable later candidates.
ADR-0003 intentionally removes fixed inference deadlines; historical
ADR-0029 descriptions of a 90-second per-probe bound are no longer current.

Acceptance

• Reuse provider validation and preserve the eight-ready target, free/ZDR
  selection, sixteen-base-probe budget per stage, and four escalations shared
  across primary/fallback.
• One or eight pending probes cannot prevent eight later responding routes
  from forming the validated serving pool within the existing probe budget.
• No pending inference is cancelled or classified as unavailable. Pending
  candidates do not enter the serving pool.
• Explicit terminal responses and unexpected worker faults remain distinct;
  all-unavailable pools fail closed, and priced fallback follows primary
  failure only.
• Record pending rows and deployment revision; focused tests, hosted
  current-head checks and independent review are required. Local validation
  alone does not prove deployment or successful live provider review.

Owner: Codex continuation of the fast-mlsirm #2114/#2120 independent review,
2026-09-27. Implementation is isolated from the active coordinator worktree.