data ingest --help copy is stale — 'Supports 9 tasks' (now 14) + MLM says texts/ but requires sequences/

[LukasWodka]

Found during functional review of the data-ingest flow on develop (2026-07-10). Two help-copy drifts in the data ingest Long text (internal/cli/data.go):

1. "Supports 9 tasks" — now 14

data ingest --help says "Supports 9 tasks (image classification, object/keypoint detection, text classification, masked language modeling, and the tabular / time-series family)" while the authoritative --task enum in the same output lists 14 tasks (adds causal_language_modeling, seq2seq, token_classification, sentence_pair_classification, embeddings — wired by #209). Self-contradictory on one screen.

• internal/cli/data.go:~149 (the "Supports 9 tasks" line) and the text layout example (~:172).

2. MLM sidecar: help says texts/, code requires sequences/

The text layout example groups "text (classification, masked language modeling)" under a single texts/ example, but masked_language_modeling's primary_subdir is sequences/ (internal/schema/layout.v1.json, read by TextSidecarDir). Following the help verbatim fails:

$ tracebloc data ingest ./reviews --task masked_language_modeling --no-input --dry-run # reviews/ has texts/
Error: missing sequences/ subdirectory in ".../reviews"

(The runtime error is self-documenting, so real-user impact is small — but the help is wrong.)

Fix

Refresh the overview count + the text-family enumeration, and split the MLM example to show sequences/.

Severity: LOW (docs). Pre-existing; #209 expanded the family without refreshing this prose. Verified on the real develop binary.

[aptracebloc]

Implemented and verified — closing.

Fixed by #233 (merged to develop 2026-07-13; body declares Closes #215, but the non-default merge target meant it didn't auto-close). Both help-copy drifts in data ingest's Long text (internal/cli/data.go) are fixed by deriving the values from source-of-truth so they can't go stale again — verified on origin/develop:

1. "Supports 9 tasks" — now Supports %d tasks rendered from len(push.SupportedCategoryIDs()) (the CLI-supported registry), so the overview count always matches the --task enum in the same output. The hardcoded "9" string is gone.
2. MLM sidecar texts/ vs sequences/ — the text-family example now interpolates push.TextSidecarDir("text_classification") and push.TextSidecarDir("masked_language_modeling"), which read primary_subdir from the vendored layout contract (layout.v1.json) — so it prints sequences/ for MLM and texts/ for classification, matching what the stager actually requires.

Guardrail: TestTextSidecarDirMirrorsContract enforces the Go registry and the vendored layout contract agree, and scripts/sync-schema.sh --check catches drift in CI.