Capture the chat session that was previously reconstructed

[ivanvyd]

Every terminal image in the documentation was real captured output — except the one showing the
feature this project is built around.

chat.svg was reconstructed from the literal strings in ChatCommand.cs, because the REPL
refuses to start without an interactive terminal and cannot be piped to a file. That was disclosed
in generate.py's docstring, a place nobody reads, while the image itself carried no caveat. In a
repository whose entire pitch is labelling what has and has not been verified, it was the one place
the labelling did not reach.

It is now a real session, captured against the live Azure workspace on 2026-08-06 by copying the
terminal buffer, with the workspace's identifiers replaced by the synthetic ones the other
transcripts already use.

What capturing it verified

Three things that had never been observed end to end, now recorded in docs/compatibility.md:

Path	Result
REPL start against a live Agent	Banner, Agent name, prompt
A question	Answer text plus a rendered result table
A follow-up — "and break that down by quarter"	Kept its context and re-grouped the same figures by quarter, without the question restating them
/sql	Printed the SQL behind the follow-up, not the first question
/exit	Clean exit

That follow-up is the behaviour README.md argues the project exists for. Until now nothing had
watched it work — the underlying API call was covered live, the loop around it was not.

ROADMAP.md listed chat as one of three v0.1 paths that resist automation, needing a human at a
terminal. That one is now struck through and dated; the other two remain.

A defect this exposed

Substituting the table name shortens the line inside a box-drawing border, so the displaced padding
has to be restored or the box renders ragged. That had already happened and shipped: ask.svg's
FROM line has been short since its own identifier substitution, in an image the README displays
today. Both are fixed, and generate.py now warns whoever does this next.

Known cosmetic limitation, stated rather than hidden

The session was captured at a 152-column terminal, so chat.svg is 1052px wide against ask.svg's
791px. It is embedded at a width that keeps it readable, but its text renders smaller than the other
images'. A narrower re-capture would match more closely — that is polish, not correction, and did
not seem worth a second live session on its own.

Verification

229 tests green. All relative links resolve. No workspace identifier appears in any transcript or
generated SVG (checked for the real catalog, schema, table, workspace host and Agent id). Regenerating
is idempotent — running generate.py twice produces no further diff.