Walkthrough re-cut: visible pointer and synced highlights on the narrated screen region

[taylormade7978]

Viewer feedback (2026-08-23, from a non-technical viewer of vforce-walkthrough.html): the narration describes parts of the screen without anything showing where to look. The viewer needs the pointer, or some way of highlighting the area being narrated, throughout the video.

This is the cold-engineer test applied to the video: a viewer who does not already know the product cannot map the words to the pixels. Requirement for the next re-cut, and standing for every future cut:

1. Visible, emphasized cursor at all times: enlarged pointer, click ripple on every click, and the cursor travels to the element BEFORE the narration names it, never after.
2. Region highlight synced to narration: while a sentence describes a screen area, that area is emphasized (outline plus surrounding dim, or a zoom-and-pan into the region). Not color-only, so the emphasis survives color-blind viewing and small screens.
3. The narration script becomes a timeline of (timestamp, screen region, sentence) so highlights are produced deterministically at capture time, not hand-edited after. If the capture is browser-scripted, inject the highlight as a CSS overlay on the narrated selector; the same triples then drive the WebVTT cues, which keeps captions, narration, and highlights in exact sync on every re-cut.
4. Highlights must not collide with the caption band below the picture.

Sequencing note (owner decision pending): hold the re-cut until vforce-lakehouse#3123 (demo tenant zero vectors) is fixed and the a4a Phase 5 purge has run, so one re-cut absorbs the weekend's platform changes plus this requirement. Cross-referenced on vforce-lakehouse#3135 (in-product walkthrough surface).

[taylormade7978]

The re-cut blocker is cleared. vforce-lakehouse#3123 is fixed and closed.

The demo tenant's document corpus was destroyed by an admin corpus purge in
mid-July: the Iceberg truncate dropped the table's metadata pointer, leaving
16,062 orphaned files on the warehouse, and the demo-reset CronJob that would
have re-seeded is disabled everywhere, so nothing put it back. Semantic search on
demo.lakehouse.vforce360.ai answered total: 0 on every query.

Live now on the public anonymous surface:

	before	after
documents list	11	4,189
Qdrant vectors	0	22,811
warm rows	0	4,189
semantic search	total: 0	real vector hits

$ curl -s -X POST https://demo.lakehouse.vforce360.ai/api/v1/search \
    -d '{"query":"bill of lading shipment","mode":"semantic","limit":5}'
  0.664  doc-bill-of-lading-5531   Straight Bill of Lading for a shipment from Northwind Traders
  0.580  doc-purchase-order-88     purchase order issued by Meridian Manufacturing
  0.504  doc-incident-report-411   incident report detailing a near-miss in a warehouse

The natural-language analyst is grounded on the restored corpus too: asked how
many pages there are per doc_type, it writes the query against
cedms.microfiche.pages and the query returns 541 doc-type groups (unclassified
1194, Report 816, Table 303, ...).

So a capture of the search and analyst segments on demo will now show results
rather than an empty state. One thing worth knowing before the re-cut: the demo
corpus is synthetic business documents and NIST/OMB publication pages, not the
Reclaim The Records genealogy corpus that lives on edgent and a4a. Queries in the
narration should be written against what this tenant actually holds; "bill of
lading", "incident report" and "cybersecurity framework tiers" all return
well-ranked hits.

Full evidence and root cause on edgentx/vforce-lakehouse#3123.