Ask the blocking questions where they arise, cut the index, and finish the button primitive

[datlechin]

The four content decisions you asked me to draft, the button migration, and a
script so the cold-start number can stop being a claim.

The two questions that stop a download now sit under the hero

"Is it really free, or free for now?" and "Can I use TablePro at work under
AGPLv3?" were items one and two of a FAQ at position fifteen — behind roughly
three thousand words, past the point where a reader holding either objection has
already left. The AGPL one blocks the highest-value visitor there is: someone
evaluating a Team licence for a company.

"Can the AI drop a table on production?" moves into the Agents section, which is
the section that raises the fear. It was being answered two positions and
several hundred words later.

"Can I move my connections from another client?" is dropped from the homepage —
SwitchFrom is an entire section on exactly that, and since #2 it sits at
position seven rather than eleven.

Nothing leaves the site. home-faqs.ts splits into homeFaqs (the four
that stay in the block) and relocatedFaqs (the four rendered inline), and
/faq spreads both — still fourteen questions, each answer existing in exactly
one place. Verified in the rendered page. Both StaleClaimsTest counts stay
true and keep meaning what they say, so neither assertion needed weakening.

DepthGrid: twelve cells to six

Three repeated a section above it, which makes cutting them deduplication rather
than taste — I verified each: Vim mode against the workbench ledger,
connection import against the whole SwitchFrom section, plugin
verification against the database grid's own lede. Three more (ER diagram,
query history, import/export) are table stakes every competitor also ships, and
"Import and export, including XLSX" already appears in the pricing comparison.

What survives is the set a reader could not assume: EXPLAIN visualised, server
dashboard, users and roles, Quick Switcher, CSV inspector, backup and restore.

The spec line goes with them — shortcuts, thresholds and defaults, genuinely
useful and genuinely documentation, and roughly 250 words standing between the
reader and the next call to action. The section's own comment called itself "an
index rather than twelve more pitches". It is now an index of six.

/faq gets an h1

It had none at all: SectionShell emits h2, so the page started its heading
tree at level two. sr-only, because the FAQ section already opens with its own
headline and two stacked headlines saying the same thing is worse than none.

Sponsors: structure only

A description slot per sponsor, and the cell becomes a column so a line can sit
under the mark. The copy is not in this PR — it is eight facts about eight
real companies and not mine to invent. A sponsor without one still renders its
mark, so filling them in is a data edit whenever you have them.

The button migration, and two hazards it had to clear first

Nineteen buttons, four padding scales, one primitive nobody imported. They all go
through it now. Neither of these would have been caught by a test here:

• Download.tsx assigns href imperatively through a ref after resolving
  the real asset URL. A primitive that swallowed the ref would have left both
  download buttons pointing at the generic releases page — a broken download, on
  the download page, silently. React 19 passes ref as an ordinary prop, so the
  primitive takes one.
• Download.tsx also swaps the two buttons' classes at runtime to make the
  detected architecture primary, and hard-coded both strings. That is the exact
  drift the primitive exists to prevent, reintroduced one layer down.
  buttonClasses() is exported and both halves now read from it.

Three keep their own element and borrow only the class list, because they are
not links: the header's dropdown trigger (aria-expanded, aria-haspopup, a
keyboard handler), the iPhone toggle (aria-controls), and the email form's
submit.

scripts/measure-cold-start.sh

You said the benchmark is measurable, so here is the thing to run. It times
launch to first window for TablePro and any competitor you name, purges the page
cache between every run, and refuses to report if it could not purge — a warm
start quoted as a cold one is precisely what this exercise exists to prevent. It
prints hardware, OS build, app version and the range alongside the median.

The point is not the number. It is that the method is runnable by anyone, which
is what turns "Under 1s" from a claim into a measurement — and whatever lands on
the page should link here.

Give me the output and I will build the EXPLAIN cost tree and the dated
methodology footnote on top of it.

Still outstanding

The hero recapture, which you are doing — that unlocks the Plate component and
the workbench two-state toggles as well. And the eight sponsor lines.

Verification

typecheck, build and pint clean; 96 tests / 1828 assertions with
REQUIRE_SSR=1. Rendered HTML checked for the objection strip, the inline AI
answer, four questions left in the FAQ block, fourteen on /faq, one h1 on
both, and the download buttons still rendering.