Narrow the chat image so it reads at the same size as the others

[ivanvyd]

The chat session was captured on a 152-column terminal, so its image came out 1052px wide against
791 for every other transcript. Scaled into a page that meant visibly smaller text than the image
directly above it — which reads as a broken asset rather than a wide one.

What changed

Reflowed to 80 columns, matching ask.txt. The prose re-wraps, both result tables and the SQL
box are rebuilt at the narrower width, and the SQL WHERE clause now breaks at a clause boundary
rather than mid-predicate. The image is 716px and embedded at its natural size, so its text
renders at the same scale as ask.svg.

Nothing was dropped

Checked rather than assumed — 43 content assertions plus an explicit row count:

• both questions, including the follow-up that keeps its context
• all four progress lines across the two turns
• both result tables, all 9 data rows (3 + 6, counted explicitly)
• the three prose bullets and both /sql /result /export hint lines
• the complete generated SQL, and /exit
• still no workspace identifiers: no lakewright, lakespeak_demo, adb-7405, or the real Agent
  name

Every box-drawing line is asserted to be exactly 80 characters — the previous version had a
one-character misalignment in the SQL box that I fixed rather than shipped.

The honesty bit

What changed is layout, so generate.py now says so. The transcript is reflowed output rather
than a byte-exact capture, and a file whose whole job is being evidence should not quietly blur
that line. The wording, figures, ordering and command sequence are exactly as the session produced
them; the line breaks are not.

It also records the two things that bite when editing one by hand: box lines within a frame must
share a width, and the longest line in the file sets the image width.

Verification

229 tests green. Regeneration is deterministic. All relative links resolve.