fix(smr): surface upload failures instead of silently succeeding

[mellox]

Problem

ficsit smr upload had several paths that swallowed errors and reported success, so a failed upload could publish a broken or incomplete version while the command still exited 0.

Changes

• Chunk upload ignored the response. The per-chunk request discarded the *http.Response entirely — HTTP status, the GraphQL errors array, and the uploadVersionPart result were all unchecked, and the body was never closed. A failed chunk (bad token, 413, a GraphQL error) was silently ignored and finalize proceeded on a broken upload. Extracted uploadChunk / sendChunkOnce, which now check the HTTP status, the GraphQL errors array, and the uploadVersionPart boolean, and always close the body.
• Per-chunk retry for transient failures. Transport errors (e.g. a dropped/corrupted TLS record, connection reset) and 5xx responses are now retried with exponential backoff via avast/retry-go (already a dependency). Permanent failures — 4xx, a GraphQL errors array, or an explicit uploadVersionPart: false — are marked retry.Unrecoverable and not retried.
• Short read / FD leak. The chunk read used a bare f.Read, which is not guaranteed to fill the buffer and could ship a zero-padded (corrupt) chunk; the file was also opened once per chunk and never closed. Now uses io.ReadFull and opens the file once with a single defer Close.
• Validate the upload file is a real .smod/.zip archive before creating a version (replaces the existing TODO).
• Exit non-zero on a failed finalize. The post-finalize state check returned nil on error, so the command exited 0 even when the server rejected the version. It now returns the error.

Testing

• Unit tests (cmd/smr/upload_test.go) cover: HTTP error status surfacing, a GraphQL errors array surfacing, explicit success, a missing success field (fail-closed), explicit false rejection, a wire-level multipart parity check, the retry path (transient 5xx retried then succeeds), the short-read fix, and archive validation (accept/reject by extension and by zip validity).
• End-to-end against the live API. A large multi-chunk upload (~30 × 10 MB) was run repeatedly. The retry layer transparently recovered from real transient failures mid-upload (a tls: bad record MAC and a connection reset on individual chunks) and completed all chunks — where the previous code would have silently dropped a chunk and finalized a corrupt version. The finalize state-check change also surfaced a real server-side finalize error that the old code swallowed as a successful (exit 0) run.

Notes

• avast/retry-go was already listed in go.mod; no new dependency added.
• go build, go vet, gofmt, and go test ./... are clean.

[Copilot]

Pull request overview

This PR hardens ficsit smr upload so chunk upload and finalize failures are surfaced (non-zero exit) instead of being silently ignored, reducing the risk of publishing incomplete/corrupt versions.

Changes:

• Refactors chunk uploading into helpers that validate HTTP status + GraphQL errors + uploadVersionPart result and always close response bodies.
• Adds per-chunk retries with exponential backoff for transient transport/5xx failures.
• Fixes chunk reading (no silent short reads), validates upload archives as real .smod/.zip, and adds unit test coverage for these paths.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
cmd/smr/upload.go	Adds validated chunk upload + retry/backoff, fail-closed response handling, robust chunk reads, archive validation, and proper error surfacing on finalize state-check.
cmd/smr/upload_test.go	Adds unit tests covering chunk upload error surfacing, retry behavior, multipart parity, short-read handling, and archive validation.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.