data ingest: collect + confirm the data schema for all tabular tasks (RFC-0002)

[LukasWodka]

RFC-0002 §8.1. Collect + confirm the data schema for all four tabular/time tasks (not silent inference).

The column-type schema is required by the contract and load-bearing downstream (DB types, per-column validation, and the model's num_feature_points at training). Today the CLI silently infers it, and inference has real gaps (leading-zero code 007→7; dates/booleans→text; 5,000-row sample cap).

• Infer a proposal, then show the column:TYPE map for the user to confirm or amend — never a silent guess. Surface the risky cases (id-like INT, empty-in-sample columns). Keep flag/CI working with an explicit acknowledgement.

Design-of-record: RFC-0002 (#174). Part of the data-ingest redesign epic.

[LukasWodka]

Re-scoped per RFC-0002 Principle 6 + Option A (cli#174 §14): the schema-inference rules (incl. the leading-zeros→INT fix) are the ingestor's source of truth — data-ingestors#349. This ticket keeps the collect + confirm UX (interface); the CLI's inference must mirror the ingestor's rules, verified by the value-level parity harness (backend#1009). Depends on data-ingestors#349.

[LukasWodka]

Correctness backbone shipped in #210 — the di#349 tabular type-inference mirror (leading-zeros→VARCHAR, INT/BIGINT, VARCHAR sized by longest value, empty→VARCHAR(1)), explicit spec.schema emit (so the CLI's correct types are authoritative in-cluster regardless of the deployed ingestor version — fixes 007→INT even pre-release), the mixed-timezone→VARCHAR fix, and a venv-free 21-case parity test pinning the CLI to di#349's committed contract. Risky columns (id-like INT, empty-in-sample) are surfaced as warnings, and --schema is a working override.

Remaining on this ticket — the interactive show → confirm/amend UX. Today the id-like/empty warnings print after the confirm prompt, so a user can't amend in-flow (they'd re-run with --schema). The follow-up: show the inferred col:TYPE before the confirm and let the user accept/amend inline.

Priority: P2 / non-blocking. Correctness is already covered by the correct inference + explicit emit + --schema, so this is ergonomics, not a correctness gate. #210 is marked Part of #185 (not Closes) so this stays tracked rather than vanishing on merge.

[LukasWodka]

FR note (2026-07-10) — functional review of develop (post-#210).

The di#349 inference engine shipped in #210 works correctly: leading-zeros → VARCHAR, Inf/Infinity → VARCHAR, ints → INT, floats → FLOAT (21/21 vendored parity cases + timezone/non-finite tests green). But the "confirm" half of this issue is not yet shipped:

• The CLI prints only a column count — Inferred schema for N column(s) from <file> (override with --schema) (internal/cli/data.go:661) — and never displays the inferred col:TYPE map on any channel (human, --verbose, or --output-json).
• There is no confirm/amend step. To override a mis-inference intelligently, a user must currently re-derive the inferred types by hand.

#210 disclosed this scope-cut (the confirm-step code was removed as dead code and the PR was scoped "Part of #185"), so this issue correctly stays OPEN. Remaining work: surface the inferred col:TYPE map — at minimum display it; ideally a confirm/amend step. internal/cli/data.go:661 is where the count is printed today.

Related follow-up filed during the same FR: #213 (the explicit --schema path isn't type-validated locally).