Support theme synthesis

Theme the records I am giving you.

First, count. How many records, over what period, from what source, and
how many you had to exclude and why. Every share you quote later is
against this denominator, stated once here.

Then the themes. Build them from the text rather than from a list you
already have. For each theme:

  Theme. A sentence describing what the customer is experiencing, not a
  one word label.
  Count and share. The share carries the denominator with it.
  Evidence. Two or three short quotes, each keyed to a record id. Strip
  names, addresses, order numbers and anything else identifying.
  Boundary. What is in this theme and what is deliberately not, so it
  does not quietly absorb its neighbours.

Then three sections.

Contradictions. Records that say the opposite of a theme. Every real
theme has some. If you found none, say you found none and treat that as
a reason for doubt rather than a strong result.

Cannot conclude. Questions these records cannot answer: how common this
is among people who did not write in, whether it is getting worse,
whether it costs anything.

Unthemed. The records that did not fit anywhere, with a count. A high
number here is information, not failure.

Rules. No sentiment score and no percentage positive. Ratings appear
as distributions with an n or not at all. Do not recommend a fix.