GSC cannot read any sitemap served from Vercel - cross-host experiment isolates the platform - Help

Since May 2026, Google Search Console has reported "Sitemap could not be read" (list view: "Couldn't fetch", 0 discovered pages) for every sitemap on two production domains I host on Vercel: dogshow.lol (project dogshow) and www.dollarbets.lol (project dollarbets). After eliminating every response-level variable, a controlled cross-host experiment isolates the serving platform itself. Posting here because Hobby-plan support routed me to the community; I believe this is a platform-level bug worth engineering's attention.

What we ruled out (each with a deploy + fresh GSC submission)

Content - the XML validates clean against sitemaps.org (no BOM, no control chars, valid element order/values). A plain-text sitemap (one URL per line) also fails, so it is not XML parsing.
1. Headers - set Content-Type: application/xml; charset=utf-8 and Content-Disposition: inline (removing the filename= param Vercel adds to static files). Still fails.
2. 1. Google-side fetch caching - submitted cache-busted, never-before-fetched URLs (sitemap.xml?v=cd2). Read fresh, still fails.
  2. 1. Static serving path - moved /sitemap.xml behind a 307 redirect to a Python serverless function returning identical bytes with explicit headers. Google read it the same day. Still fails.
    2. 1. Compression - not the discriminator; the working non-Vercel host (below) also serves brotli to br-accepting clients.
    3. Meanwhile, on the same domains: GSC URL Inspection live test on the sitemap URL itself returns "URL is available to Google", and regular pages crawl and index normally. Only the sitemap-processing pipeline fails, and only on the Vercel-served hosts.
  3. The decisive experiment (23 Aug): I proxied the byte-identical sitemaps through a non-Vercel host (a Node service on Render) and submitted them in the same Search Console account:
3. - headshotswithabird.com/dogshow-sitemap.xml -> Success, 89 pages discovered, read within minutes
  - - headshotswithabird.com/dollarbets-sitemap.xml -> Success, 110 pages discovered, read within minutes
  - Same bytes, same GSC account, same day. The only variable changed was the serving host. Months of failure on Vercel; instant success off it.
4. Conclusion: something in the interaction between Vercel's edge and Google's sitemap fetcher specifically (which evidently behaves differently from Googlebot's page fetcher) breaks sitemap retrieval at a level below response headers, compression, and static-vs-function serving - plausibly connection/TLS/protocol-level handling of that particular client. I can't see below the response layer from my side; Vercel can.
Happy to provide project IDs, deployment IDs, and exact GSC read timestamps, and to run further controlled tests on my projects. Has anyone else seen GSC sitemap failures on Vercel-hosted sites that page-level crawling doesn't explain?