Link-health triage for data/sources/ on 2026-10-04 - live/redirected/blocked counts and the two-probe rule that kept 70 URLs out of the dead list

METHOD AND NEGATIVE RESULT, url is a representative live source endpoint. COUNTS OVER 1813 source records / 1507 distinct URLs, probe date 2026-10-04. Records: 1623 live, 103 redirected-and-resolving, 78 indeterminate-blocked, 9 dead. Distinct URLs: 1340 live, 88 redirected-and-resolving, 70 indeterminate-blocked, 9 dead. RETRIEVAL-DATE DISCIPLINE, ALREADY SATISFIED: every record in data/sources/ already carries a valid schema-conformant 'accessed' date (2026-10-03 on 1120 records, 2026-10-04 on 693), and 'updated' equals 'accessed' on all 1813. No record needed a date added. 1652 records have archived_url null, which is correct rather than a gap: most have no archive, and inventing one would be a lie. THE 70 BLOCKED URLs ARE NOT DEAD AND MUST NOT BE COUNTED AS DEAD: docs.vllm.ai (19, HTTP 429 rate-limit), support.huawei.com (5), bhphotovideo (4), overclockers.co.uk (4), news.ycombinator.com (3, HTTP 419), intel.com (3, 403 'Access Denied'), microcenter.com (3), doi.org (4, HTTP 202 = accepted-then-challenged), plus singletons on geizhals, huggingface, reuters, sec.gov, supermicro, dl.acm.org, w3.org, emp.lbl.gov, openai.com and four read-timeouts on marketplace.nvidia.com / bestbuy / hpe. AN EARLIER PASS HAD ALSO RETURNED 404 FOR THE FIVE github.com/mlcommons/* REPOSITORY ROOTS (src5-mlperf-inference-github, src5-mlperf-inference-results-v3-1, -v4-1, -v5, and related): an independent GitHub contents-API check plus a browser-UA GET returned 200 for all of them, so that was a transient probe artefact, NOT five dead repos, and no note was written claiming otherwise. That is the second false negative this audit caught in its own instrument, after the Cerebras one the repo already learned from, and it is the reason DEAD requires two independent probe failures plus a content-level check of the response body. REDIRECTS ARE NOT DEAD AND NOT INVISIBLE: 88 distinct URLs redirect and then resolve, and several are substantive moves worth knowing about - cloud.google.com/compute/gpus-pricing -> /products/compute/pricing#gpu-attachment-pricing, cloud.google.com/tpu/docs/* -> docs.cloud.google.com/tpu/docs/*, aws.amazon.com/machine-learning/{inferentia,trainium} -> /ai/machine-learning/*, blog.vllm.ai -> vllm.ai/blog/*, cdrdv2.intel.com/v1/dl/getContent/817486 -> cdrdv2-public.intel.com/817486/gaudi-3-ai-accelerator-white-paper.pdf (Mellanox doc folded into NVIDIA's CDN), azure pricing/details/storage/disks -> /managed-disks/. A redirect also FAILED to save one record: src5-infiniband-docs redirects mellanox.com to network.nvidia.com and the redirect TARGET is 404, so checking only the pre-redirect status would have scored it live.