[NamelsKing]

Ordinary authenticated upload/link failures could lose their response when the server immediately destroyed the stream, leaving CI with only a connection reset. This delivers written failure records before closing and reports the stable reason and request phase in CLI output and JSON.

• Preserve remote failures received while uploading, without replacing local source errors or accepting an early success.
• Report unwrapped upload filesystem failures with existing codes and unexpected failures as SERVER_ERROR; preserve explicit storage-guard codes. Logs add a fixed phase and allowlisted errno without paths, messages, stacks or hook secrets.
• Bound repeated unmanaged-path warnings to 32 names plus a count; unmanaged paths remain untouched.
• Bump to 0.6.3 and document compatibility, failure output and rollout.

Validation: complete Node 22/24 and Bare suites passed with 622 tests each before the wording correction; all 134 affected Node 22 tests passed after correction. Final lint/public types and package checks passed. A real local DHT fixture using the published 0.3.0 client verified file/directory uploads, unchanged retries and fail-closed handling of the new code. Independent review found three wording issues; all were corrected and re-reviewed clean.

Both client and server should move to 0.6.3 for the improved diagnostics. This changes no upload limits, reservation accounting, protocol fields, authentication, hooks or retention policy, and does not solve a real lack of disk space. Older clients still upload successfully but reject the new SERVER_ERROR code as an invalid record.