Preserve semantic text represented only inside captured Wix offer-card media

## Problem

The visual hero clipping reported in Automattic/static-site-importer#1389 is fixed on current trunk, but the offer card's meaningful text remains absent from Gutenberg markup and the frontend DOM:

- `Your first two visits`
- `NEW PATIENT SPECIAL`
- `GRAND OPENING`

The card looks visually complete only because it is preserved as raster media. The text therefore survives as pixels but is inaccessible, unsearchable, and uneditable.

Current-trunk geometry is healthy: the 111px header is followed by a fully visible eyebrow at y=294 and complete headline at y=327..446. This tracker is only the remaining semantic-content loss.

## Required investigation

Determine whether the captured artifact carries these strings in source HTML, semantic evidence, media metadata, accessibility labels, or associated responsive variants. When textual evidence exists, preserve it through a generic accessible/native Gutenberg representation without visually duplicating the rasterized card. If the strings exist only in image pixels, identify the owning capture-producer gap rather than adding OCR or site-specific text.

## Evidence

Parent report: Automattic/static-site-importer#1389. Verified on the same captured `360chiro.co.uk` artifact with current Blocks Engine trunk (`0.8.1` source), current SSI main, and WordPress 7.1.

## AI assistance

GPT-5.6 Sol via OpenCode recompiled and materialized the source artifact, separated the fixed geometry from the remaining DOM/content loss, verified the rendered card is raster-preserved, and drafted this upstream tracker. Chris Huber directed the work and remains responsible for it.

Start by inspecting the captured 360chiro.co.uk artifact and its source HTML, semantic evidence, media metadata, accessibility labels, and responsive variants, then trace how Blocks Engine 0.8.1 and current SSI materialize the offer card. Done means preserving textual evidence in a generic accessible/native Gutenberg representation without duplicating raster media, or documenting the capture-producer gap if the strings exist only as pixels.

Written by the indexing model from the issue text.