data ingest: mis-cased media folder + labels.csv sniffs as confident tabular → silent mis-ingestion

[saadqbal]

Found while reviewing #198 / following up on #180b.

Problem

When a dataset directory has labels.csv plus a mis-cased media folder (e.g. Images/, Texts/, Sequences/ instead of the literal images//texts//sequences/), SniffFamily (internal/push/preview.go) reports a confident tabular layout:

• the subdir switch is case-sensitive, so Images ≠ images → hasImages stays false;
• labels.csv counts toward csvCount (== 1);
• so it hits the !hasImages && !hasText && csvCount == 1 branch → confident tabular, echoes "Found a CSV table — this is tabular data."

The guided data-first flow then skips the family question (resolveFamily returns on Confident) and offers only tabular tasks. The user picks a tabular task, DiscoverTabular runs, findSingleCSV ignores subdirectories and finds the one CSV (labels.csv) → it succeeds. Result: an image/text dataset is silently ingested as a standalone table, the media files dropped, no error.

Why the existing test misses it

TestSniffFamily has a mis-cased Images/ is not confident image case — but it only asserts the sniff isn't confident image. It never checks it isn't confidently tabular, which is the actual footgun.

Proposed fix

When the lone CSV sits next to a subdirectory whose name matches a marker case-insensitively but not exactly (a likely mis-cased media folder), stay ambiguous and ask the family question instead of confidently claiming tabular. Narrow — an unrelated subdir (backup/, raw/) still sniffs confident tabular, matching DiscoverTabular, which ignores it.

Acceptance

• labels.csv + Images/ (or Texts//Sequences/) → sniff NOT confident (asks family).
• labels.csv + images/ (exact) → still confident image.
• single CSV + unrelated subdir → still confident tabular.
• Regression test pinning the mis-cased-tabular case.

Rolls out of the #180 / #198 line.

[LukasWodka]

Taking this over from @saadqbal — building the fix now. Approach per your proposal: when the lone CSV sits next to a subdir whose name matches a media marker case-insensitively but not exactly (a likely mis-cased images/texts/sequences folder), stay ambiguous and ask the family question instead of confidently claiming tabular. An unrelated subdir still sniffs confident tabular, matching DiscoverTabular. Thanks for catching it in the #198 review.