feat(dataset push): interactive guided mode

[aptracebloc]

Goal

Let a user run tracebloc dataset push and be prompted for what's missing instead of needing every flag up front — the streamlined guided flow. Flags still win; non-TTY/CI behavior is unchanged.

Scope

• When required inputs are missing AND stdout is a TTY AND --no-input is not set: launch a guided flow using AlecAivazis/survey/v2:
  • path → category (select from the supported list) → table (validated via push.ValidateTableName) → intent (train/test select) → label-column → category-specific fields:
    • image: target-size (default = auto-detected from the first image),
    • tabular/time-series: --schema preview from inferred types, label policy,
    • keypoint: number-of-keypoints,
    • text/MLM: confirm the texts//sequences/ (+ tokenizer.json) layout.
  • A confirm screen showing the assembled spec (reuse Phase 1–2 rendering) before running.
• Prompts populate the existing SpecArgs — the downstream synth → validate → stage → submit path is unchanged.
• Graceful degradation: non-TTY or --no-input → today's flag-only behavior + the same clear errors.
• Suggested split: PR-a = prompt engine + path/category/table/intent/label; PR-b = category-specific dynamic prompts + confirm screen (mirrors the chore(data ingest): re-land #147 preflight parity onto ux (stacked-base miss) #151 PR-a/PR-b pattern).

Acceptance criteria

• Missing-flag + TTY → guided prompts; all flags present → no prompts (unchanged).
• --no-input / non-TTY → no prompts; missing required → same exit-2/3 errors as today.
• Prompted values flow through the existing SpecArgs.Build + schema validation, unchanged.
• Confirm screen; declining aborts cleanly (no cluster mutation).
• Prompt-mapping logic unit-tested (scripted / in-memory survey transport); non-TTY path asserted.
• make ci green.

Go concepts we'll cover

• Interactive prompts (new dependency — survey/v2): select vs input vs confirm, reading/validating stdin.
• Layering an optional interactive front-end over an existing flag-driven flow without duplicating logic.
• Testing interactive code (injecting a non-terminal transport / forcing --no-input).

Notes

• Decision locked: survey/v2 (not huh/bubbletea). Depends on #PHASE1 and #PHASE2 for the confirm/summary look. Branch off develop.

Part of the UX/interactive epic.

🤖 Generated with Claude Code