data ingest: task-irrelevant flags are silently accepted and ignored (e.g. --target-size on tabular, --schema on images)

[saadqbal]

Severity: LOW (UX/footgun). Found during the #67 stress sweep; framing refreshed 2026-07-11 to the current data ingest / --task verbs (RFC-0002 renamed dataset push --category → data ingest --task). The bug itself is unchanged and still live.

Repro (all pass pre-flight, flag silently ignored)

tracebloc data ingest ./tab/... --task tabular_classification --target-size 512x512      # target-size meaningless for tabular
tracebloc data ingest ./tab/... --task tabular_classification --time-column created_at    # time-column only for time_to_event_prediction
tracebloc data ingest ./tab/... --task tabular_classification --number-of-keypoints 17    # keypoints only for keypoint_detection
tracebloc data ingest ./img/... --task image_classification  --schema a:INT               # schema only for tabular / time-series

Each runs as if the flag weren't there — no error, no warning.

Current state (verified 2026-07-11)

All five task-scoped flags are still registered unconditionally in internal/cli/data.go (--target-size ~L308, --schema ~L316, --label-policy ~L319, --time-column ~L322, --number-of-keypoints ~L324). There is no flag-relevance guard anywhere — the only Flags().Changed uses in data.go are for the --table/--category alias resolution, not for relevance checking. So an out-of-task flag is silently dropped.

Why it matters

A user who sets --target-size 256x256 for a tabular ingest (or --schema for images) reasonably expects it to take effect. Silent acceptance means "I set it but it did nothing" with no feedback — the classic footgun. --help documents which flags belong to which task, but nothing enforces or warns at runtime.

Expected

When a flag that's only meaningful for task X is set with task Y, emit a warning (Note: --target-size is ignored for tabular_classification) — or, for clearly-nonsensical combinations, error. At minimum warn; don't silently drop. A single taskScopedFlags relevance table checked after task resolution covers all five.

Part of #67.

[LukasWodka]

🔎 Refinement deep-dive — verified on develop a2c5816 (stress-verified)

Still valid: confirmed reproduces. CategorySpec (category.go:19) has no relevance field; each category-scoped flag is read only inside its family branch and dropped otherwise, with no Flags().Changed guard — e.g. data ingest --category tabular_classification --target-size 512x512 never reaches ParseTargetSize (it's in the IsImage branch), the value is silently discarded, and the run proceeds to cluster discovery. No warning, no error.

Fix (size S, independently shippable):

• category.go — single source of truth, derived from existing predicates (not 16 hand-listed arrays): CategoryScopedFlags = {target-size, schema, label-policy, time-column, number-of-keypoints} + FlagRelevant(category, flag) bool (target-size→IsImage; schema→IsTabular; label-policy→IsRegressionClass; time-column→time_to_event_prediction; number-of-keypoints→keypoint_detection; default→true).
• data.go — capture explicitly-set scoped flags via Flags().Changed (same mechanism as CategorySet); add a guard after the category gate + interactive resolution (so the resolved category is final): any set flag where !FlagRelevant → exit 2 naming the flag and the categories it does apply to. Matches the sibling exit-2 flag guards (--overwrite+--idempotency-key, unsupported-category). Warn-vs-error: recommend error; flag the choice to Asad (issue owner) in the PR — the softer Warnf uses the same seam.

⚠️ Stress-verification added one fix: guard must skip when category is empty (if a.Spec.Category != "") — otherwise --category="" --target-size=… fires a misleading "not relevant for ''" and preempts the canonical "category is required" from schema validation.

Tests: TestDataIngest_IrrelevantFlag_ExitsTwo (table: tabular+target-size/time-column/keypoints, image+schema/label-policy → exit 2; fails today, falls through to exit 3) plus a negative over-fire case (image+--target-size=4x4 must reach the bad-kubeconfig exit 3, not 2 — load-bearing). TestFlagRelevant in internal/push.

Cross-issue: the relevance table is the same per-category knowledge #105's dispatch needs — keep it in category.go as the single source so they don't drift. Shares the data.go RunE/args region with #75/#76 (merge-order coordination). --label-column for masked_language_modeling is the same bug class but a core prompted flag — deliberately scoped out (interactive false-positive risk).

Ready for refinement.