feat(onboarding): Confirm-and-refine derived understanding before generating the planner

Size: M — improves onboarding understanding with or without images.

Why

Today the flow jumps straight from a short chat to generating a full planner once isProfileSufficient() heuristics pass (conversation.ts:30-40: ≥4 facets, ≥3 dimensions, one salient). The user never sees or corrects what we inferred, so wrong/low-confidence facets (and, once image attachments land, anything mis-read from a photo) silently shape the generated sections. A lightweight confirmation step measurably improves fit and trust.

Scope

1. Before /api/profile/generate, show a compact "Here's what I understand about you" summary: facets grouped by dimension with salience, plus the sections we're about to create.
2. Let the user tweak: remove/soften a facet, add a missing one, correct an image-derived detail, rename/deselect a proposed section.
3. Feed edits back into the profile (reuse applyFacets) so generation and future reconciliation start from a corrected baseline.
4. Flag low-confidence / image-derived facets distinctly so they're easy to verify.

Notes

• Keep it skippable ("Looks good, build it") to preserve the fast path.
• Naturally reuses the profile-editor and reconciliation-review surfaces once those exist — this is the onboarding-time entry point to the same idea.

Start with conversation.ts:30-40 and trace the flow before /api/profile/generate. Read the existing profile-editor and reconciliation-review surfaces, plus applyFacets, to understand how edits can be reused. Done means users can review and edit facets and proposed sections, identify low-confidence or image-derived details, and skip confirmation with “Looks good, build it”.