data ingest/delete: misleading validation error messages (delete '(set --name)' on a positional arg, keypoints 0/neg, --label-policy type-noise)

[saadqbal]

Severity: LOW (error-message quality). Three confusing messages found during the #67 stress sweep; framing refreshed 2026-07-11 to the current data ingest / data delete verbs. All three still reproduce.

(a) data delete "" points at a flag the command doesn't have

tracebloc data delete ""
# → Error: ... dataset name is required (set --name)

But data delete takes a positional <dataset> argument — there is no --name flag on delete (that's data ingest). The message comes from the shared push.ValidateTableName (internal/push/spec.go ~L100-102), and data_delete.go (~L110) calls it with no empty-arg pre-guard. Should be: dataset name is required — pass it as an argument: tracebloc data delete <dataset>.

(b) --number-of-keypoints 0 (or negative) says you didn't provide it

tracebloc data ingest ... --task keypoint_detection --number-of-keypoints 0
# → Error: keypoint_detection requires --number-of-keypoints ...
tracebloc data ingest ... --task keypoint_detection --number-of-keypoints -5
# → same "requires" message

The user did pass the flag. data.go (~L719) branches on NumberOfKeypoints <= 0 with a single "requires" message and no Changed/set-tracking, so it conflates "unset" (Go int zero value) with "set to an invalid value (≤ 0)". Distinguish them: --number-of-keypoints must be a positive integer (got 0).

(c) Bad --label-policy leaks an internal schema representation

tracebloc data ingest ... --task tabular_regression --label-policy banana
# → synthesized spec failed schema validation (2 issues):
# label: got object, want string <-- internal noise, baffling to a user
# label.policy: value must be one of 'passthrough', 'bucket'

The second line is the real, useful error. The first ("label: got object, want string") is an artifact of how the regression label is encoded as an object and should be suppressed when the more specific label.policy error is present (no type-noise suppression exists in validate.go today).

Part of #67.

[LukasWodka]

🔎 Refinement deep-dive — verified on develop a2c5816

Still valid: all three sub-bugs reproduce. Three small, independent edits.

(a) data delete "" says "(set --table)" but delete takes a positional arg. spec.go:77 (ValidateTableName) hardcodes the flag-caller's message; data_delete.go:110 (positional <table>, no --table) reuses it.
→ At data_delete.go before the ValidateTableName call: if a.Table == "" { return exit-2 "table name is required — pass it as an argument: tracebloc data delete <table>" }. Shared validator untouched (ingest/interactive callers unaffected; spec_test.go untouched).

(b) keypoints conflates unset(0) vs invalid(neg). data.go:548 NumberOfKeypoints <= 0 gives one "requires --number-of-keypoints" for both.
→ Thread NumberOfKeypointsSet bool via cmd.Flags().Changed("number-of-keypoints") (mirrors the existing CategorySet precedent at data.go:221); split into unset → "requires…" vs <=0 → "--number-of-keypoints must be a positive integer (got %d)". Interactive path sets a positive value first, so neither branch fires there.

(c) schema prints a noisy label: got object, want string alongside the real label.policy enum error (label is oneOf[string,object]; --label-policy banana trips both branches; validate.go flattens all leaves).
→ In validate.go: add unexported isType bool to ValidationError (set via a *kind.Type check — robust, no message matching), then after flattenValidationError drop any leaf that is a type error and a strict ancestor (by dotted path) of another leaf. Kept out of FormatErrors (stays a pure formatter per its pinned tests).

Tests: TestDataDelete_EmptyTableName_ExitsTwo; TestDataIngest_Keypoints_UnsetVsInvalid (0/-5 cases fail today); TestValidate_SuppressesOneOfTypeNoise + keep the "regression w/o explicit policy" case green to prove no over-suppression.

Size S, independently shippable. Over-suppression in (c) is bounded (needs isType and a strict-descendant leaf — only the oneOf split). Cross-issue: (b) lives in the per-category block that #77/#105 may later centralize.

Ready for refinement.

[aptracebloc]

Implemented and verified — closing.

All three sub-bugs were fixed by #233 (merged 2026-07-13, into develop; body declares Closes #76 but the non-default merge target meant it didn't auto-close). Verified live in the tree:

• (a) data delete "" — empty-arg guard now says dataset name is required — pass it as an argument: tracebloc data delete <dataset> (no phantom flag). internal/cli/data_delete.go
• (b) --number-of-keypoints 0/negative — now distinguishes unset (requires --number-of-keypoints) from set-but-invalid (must be a positive integer (got N)) via cmd.Flags().Changed. internal/cli/data.go
• (c) bad --label-policy — suppressOneOfTypeNoise drops the parent got object, want string line when the specific label.policy enum error is present. internal/schema/validate.go

Tests added: TestSuppressOneOfTypeNoise + the keypoints unset-vs-invalid and empty-name cases.