feat(data ingest): rename --category/--table/--intent → --task/--name/--split (#180a)

[LukasWodka]

Stacked on #179 — merge after #179. Base is feat/179-plain-language-ceremony, not develop; the diff shown against develop until #179 lands will include #179's commit. Retarget to develop once #179 merges.

Part of #180. This is #180a, the mechanical flag renames from RFC-0002 §6. It does notclose#180 — the data-first flow inversion + task-picker rework is #180b, stacked on top of this.

Summary

Renames the three core data ingest flags to plain, task-first names, and keeps every old flag as a hidden deprecated alias so existing scripts don't break. The wire/spec field names (category / table / intent) are unchanged — this is a CLI-surface rename only.

Old (now hidden alias)	New (canonical)	Notes
--category	--task	old image_classification default dropped
--table	--name	wire field stays table
--intent	--split	default train

What changed

• --category → --task, and the image_classification default is dropped. Omitting the task now drives the interactive picker (on a TTY) or returns a clear which task is this data for? pass --task … error (non-interactive) — never a silent image assumption.
• --table → --name. Plain-language help ("a name for this dataset — you'll reference it by this name when you start a training run"); the k8s/PVC detail is gone from the flag help.
• --intent → --split, default train. The guided prompt is reworded to "Is this training or test data?". The default is applied after the interactive block, so guided runs still ask; non-interactive runs that omit it get train.
• Reworded the interactive prompts, the Review labels, the pre-flight Ingest settings labels, all help/error copy, and the README example to the new names.
• Hidden aliases are registered with MarkHidden (absent from --help) and reconciled in RunE: the canonical flag wins; a legacy alias fills in only when the new flag wasn't passed.

Behavior is otherwise identical — exit codes, guards, and the destination-exists (exit 6) check are unchanged. No overwrite-promptability (that's #180b).

Test plan

go build ./..., go test ./..., gofmt -s -l . (empty), go vet ./..., plus errcheck / ineffassign / misspell — all green locally.

New / updated unit tests:

• TestDataIngest_DeprecatedFlagAliases — old names resolve; a bad value via --category still reaches the task gate (aliased value flows through, not ignored).
• TestDataIngest_OmitTask_NonInteractive_Errors — omitting --task off a TTY exits 2 with a message naming --task.
• TestDataIngest_OmittedSplit_DefaultsToTrain — omitting --split no longer rejects; falls through to the injected bad kubeconfig (exit 3).
• Existing exit-code + interactive tests converted to the canonical flags; prompt-answer keys updated to the new labels.

Manual smoke test against a bad kubeconfig (built binary):

• new flags → exit 3 (spec valid, reached cluster step);
• old flags via alias → exit 3 (identical);
• omit --task → exit 2 with the clear --task error;
• omit --split → exit 3 (default train applied);
• canonical wins over alias on conflict (bad alias value ignored → exit 3);
• deprecated aliases absent from data ingest --help.

Decisions to ratify

1. --split default train is applied post-interactive (an if a.Spec.Intent == "" { … = "train" } fallback), not baked into the cobra flag default. This keeps the guided flow asking "training or test?" while a non-interactive omit still gets train. RFC-0002 §5.
2. Canonical-wins-on-conflict: if both --task and --category (etc.) are passed, the new flag wins and the alias is ignored. Documented and tested.
3. JSON --output-json keys are left as category/table/intent (wire format — renaming would break scripts, the opposite of the alias promise).
4. Deferred copy nit (not this PR): the exit-6/exit-7 error strings still say "uploading everything" — banned per tracebloc voice ("ingest", never "upload"). Left for a copy pass so #180a stays scoped to renames; only the --table→--name token was touched on those lines.

---

Note

Medium Risk
CLI flag and validation behavior changes can break scripts that relied on --category defaulting to images or numeric-leading --table names; hidden aliases mitigate rename risk but not the stricter name rule or missing-task error.

Overview
Renames the primary data ingest flags to --name and --task (plain-language help, pre-flight labels name / task), with hidden --table / --category aliases so scripts keep working; canonical flags win on conflict. Wire/spec fields stay table / category / intent; --output-json keys are unchanged.

--task no longer defaults to image_classification—omitting it triggers the interactive picker on a TTY or a clear exit-2 error off-TTY. --intent is still the flag name but now defaults to train after guided prompts (guided flow still asks train vs test).

User-facing copy shifts from “category/table” to “task/name” across help, errors, and interactive prompts (TaskSet replaces CategorySet).

ValidateTableName now requires [A-Za-z_][A-Za-z0-9_]*, aligned with the ingestor—leading-digit names like 123 fail at the CLI instead of after staging.

Tests cover deprecated aliases, omitted --task, and default intent; README example uses the new flag names.

^{Reviewed by Cursor Bugbot for commit b4e5678. Bugbot is set up for automated code reviews on this repo. Configure here.}

[LukasWodka]

Pushed the two review fixes (ba01f9e).

A — reverted the --intent→--split rename.--intent stays the canonical flag; --split is gone entirely (no hidden alias for it). The --category→--task and --table→--name renames are untouched, still with their hidden deprecated aliases. Dropped the splitVal reconciliation block and the extra flag registration, so the surface is just --intent. The wire/spec field stays intent, and the local summary + interactive review now label the field "intent". Renamed the test to TestDataIngest_OmittedIntent_DefaultsToTrain and swapped every --split=train for --intent=train.

Default-train decision (kept): omitting --intent still defaults to train. It's applied after the interactive block, so the guided flow still asks "training or test?" and a non-interactive run that omits it gets train without erroring (RFC-0002 §5).

B — tightened table-name validation to mirror the ingestor. The ingestor's validators/table_name_validator.py is the source of truth and requires ^[a-zA-Z_][a-zA-Z0-9_]*$ — a name must start with a letter or underscore. The CLI's tableNamePattern was looser (^[A-Za-z0-9_]+$), so it accepted leading-digit / all-digit names like 123 and 1data that the cluster then rejects post-upload. Changed the regex to ^[A-Za-z_][A-Za-z0-9_]*$, updated the doc comment and error message to state the leading letter/underscore requirement and that it mirrors the ingestor, and added TestValidateTableName_LeadingDigit (rejects 1data/123; accepts _data/Data1/chest_xrays_train). Also dropped the now-invalid 9starts_with_digit case from the accepts test.

go build ./..., go test ./... (all green), gofmt -l . (clean), and go vet ./... all pass locally.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit ba01f9e. Configure here.}