fix: stop the moderation log asserting an age nobody assessed

The vision prompt told the model to write the phrase "18-year-old" into
described_image_content whenever a female was detected, regardless of what
it saw. That field is labelled "objective description of the image for
moderation logs" and is stored in the nsfw_images record, so every stored
row carried an age claim produced by instruction rather than observation.
A log that says what it was told to say is not evidence, and it reads
exactly like evidence. Removed.

enforceAdult is untouched and still rewrites the generation prompt. That
one is defensible: it instructs a generator what to draw, and telling it to
draw an adult is a safety measure, not a record of anything.

Two related changes to the same classifier:

The prompt said "isMinor: true if the person appears under 18. When
uncertain, answer true." Instructing a model to convert doubt into a
finding produces noise, not caution - the same photograph was cleared on
one run and flagged on the next with nothing changed. It now asks for an
honest age estimate and sets isMinor only where something visible supports
it, named in a new minorIndicators field. The genuinely uncertain case is
still a block, and says so, because that case is real and different from
being unsure whether somebody is 22 or 26.

blockReason was the fixed string 'suspected minor present'. It now carries
the model's own indicators, so a skipped image can be told apart from a
shrug without reading raw execution JSON - which is how this whole thread
started.

The block itself is unchanged and was verified after: isMinor or an
apparent age under 18 blocks, a blocked image is never stored and never
generated, and the description and image bytes are still blanked on a
block.