docs: revalidate and update the celld integration against celld 0.5.0

[sandros94]

The celld page was written against celld 0.2.0, and since then celld had a few releases with breaking changes (latest is 0.5.0). I re-ran the whole setup against 0.5.0 and updated the page to match what actually happens on the wire now.

What changed on the page

1. celld diagnose no longer "only checks access": both diagnose and every node at startup now send four conditional writes and a ranged read under probe/, and mountx/s3 passes them. The old known problem is gone, and ranged reads are listed as the fourth bucket requirement
2. the fencing.md link was dead, celld moved that content to guarantees.md
3. the request list was outdated: celld now also uses DeleteObjects, multipart uploads (LTX segments ≥ 5 MiB and large R2 writes) and delimiter/max-keys/continuation-token on lists
4. "What works" now covers paged restore, SIGTERM handoff, crash recovery, Workers KV, the R2 binding requests and celld cell list
5. new known problem: R2 bindings lose object metadata (httpMetadata, customMetadata, checksums), since the gateway only stores x-amz-meta-mtime
6. the half-written PUT section stays, but notes that 0.5.0 did not show the failure (two nodes, 200 raced cells, zero failed requests) while the gateway behavior is unchanged. So to me this is still a ticking bomb

test/integrations/celld.test.ts only gets comment updates, as the related issue has more to it other than celld integration itself. The assertions are untouched and still pass on 0.5.0.

How it was tested

• official celld 0.5.0 release binary + esbuild, one host, one gateway process, node-fs driver over an ext4 directory
• Linux 6.18 only, which is now listed under "Not tested" since older kernels are still around in LTS distros
• ofc pnpm lint, pnpm typecheck and the celld suite (3/3)
• outside the suite, a logging gateway plus scripted scenarios for everything listed under "What works"

The R2 metadata limitation will get its own issue, as I'd rather discuss the approach there first, but first I'll comment some more on #20 with new findings and hopefully a followup proposal

[vercel]

@sandros94 is attempting to deploy a commit to the unjs Team on Vercel.

A member of the Team first needs to authorize it.