QA: adversarial stress test of the ingestion flow (dataset push/list/rm) on dev

[saadqbal]

Tracking ticket for an adversarial QA session against the CLI built from develop (2663bf3), run against the `tb-client-dev-templates` cluster (namespace `tracebloc-templates`, chart 1.6.1).

Goal: act like a real customer who does NOT follow the happy path — bad flags, malformed CSVs, broken folder layouts, interrupts, weird table names, non-TTY use, permission errors — and file a ticket per reproducible bug/UX failure found.

Method: black-box first (only `--help` and docs as a user would see), source consulted only to confirm root cause when filing.

Findings are filed as separate issues and linked back here.

[saadqbal]

Stress sweep results — 10 findings filed

Black-box adversarial run of the CLI built from develop (2663bf3) against tb-client-dev-templates (ns tracebloc-templates, chart 1.6.1, ingestor 0.3). Acted as a non-happy-path user: bad flags, malformed CSVs, broken image layouts, interrupts, re-runs, hostile identifiers. All children added to project #2.

Release-blockers

• dataset push: image ingestion only works for .jpeg — .png/.jpg/.webp fail the File Type Validator #68 [CRITICAL] image ingestion only works for .jpeg — .png/.jpg/.webp fail the File Type Validator (CLI never sends the extension field; ingestor defaults to .jpeg). Breaks the default category end-to-end.
• dataset push: --label-column is never checked against the CSV header — wrong name writes orphaned rows then fails registration #69 [HIGH] --label-column never checked against the CSV header — a typo passes dry-run and all ingestor validators, writes orphaned rows to MySQL, then fails registration.
• dataset push: re-pushing to an existing table silently duplicates data (no warning, no --overwrite/--append) #70 [HIGH] re-pushing to an existing table silently duplicates data (6→12 rows), no warning, no --overwrite/--append.

Medium

• dataset push: UTF-8 BOM not stripped from CSV header — Excel-exported CSVs get a corrupt first column name #71 UTF-8 BOM not stripped from CSV header → corrupt first column name (breaks Excel exports).
• dataset push: image pre-flight is too shallow — missing-image refs, undecodable/zero-byte images, and mixed resolutions all pass dry-run #72 image pre-flight too shallow: missing-image refs, undecodable/zero-byte images, and mixed resolutions all pass dry-run.
• dataset push: tabular pre-flight silently collapses duplicate header columns and accepts zero-data-row CSVs #73 tabular pre-flight: duplicate header columns silently collapsed; zero-data-row CSV passes.

Low / polish

• dataset push: --category help lists 5 of 9 categories, and unknown-vs-empty category errors give two different lists (one leaks unsupported segmentation) #74 --category help lists 5 of 9; unknown-vs-empty category errors give two different lists (empty leaks unsupported segmentation).
• CLI: a mistyped subcommand (e.g. dataset pus) prints help and exits 0 instead of erroring #75 mistyped subcommand (dataset pus) prints help and exits 0.
• dataset push/rm: several misleading validation error messages (rm "(set --table)", keypoints 0/neg, --label-policy noise) #76 misleading error messages (rm "" → "(set --table)"; --number-of-keypoints 0 → "requires"; --label-policy leaks "label: got object, want string").
• dataset push: category-irrelevant flags are silently accepted and ignored (e.g. --target-size on tabular, --schema on images) #77 category-irrelevant flags silently accepted/ignored (e.g. --target-size on tabular, --schema on images).

What held up well (no ticket needed)

• Table-name validation is robust and security-conscious — SQL-injection, path-traversal, slashes/dots/spaces/emoji, and >63 chars all rejected with clear, educational errors; symmetric across push/rm.
• ingest validate handles empty / binary / directory / missing / wrong-type / YAML-bomb / 95 MB inputs cleanly (exit 3, bounded memory).
• SIGINT during the job watch detaches gracefully ("Ingestion continues… reconnect with kubectl logs -f"), no orphaned stage Pod/Job.
• Bad --context/--namespace/--kubeconfig give clear, actionable errors.
• --output-json emits a clean structured error envelope on failure.
• Tabular happy path and dataset list/rm (incl. recovery of orphaned/failed-ingest tables) work well.

Cluster left clean — all qa_* test tables created during the sweep were removed with dataset rm.