feat(loop): the digest judges intent from the content, not the file type

[wu908]

The intent on a digest deliverable came from the input medium:

INTENT_BY_KIND = {"text": "solve", "link": "solve", "idea": "share",
                  "image": "record", "voice": "record"}

So a piece of text that tells a story was classified 解决, and an image captioning a process was classified 记录. The creator never sees this happen — they see the consequences: the system asks a solve-flavoured question and watches solve-flavoured signals for a piece that is not one. Same class as the 展示作品被问转折 bug, different root: not too few buckets, but the wrong evidence for the bucket.

What changed

The model already reads the whole material, so it now judges it too:

• the judgment contract gains content_intent (one of the three, or null);
• the prompt says to judge by content, and explicitly not to pick solve just because the material is text;
• the JSON example carries the field (a described-but-not-exemplified field is the shape behind bug fix(loop): make the digest AI draft actually parse #85);
• when there is no model — or it declines to judge — the medium mapping still applies, unchanged.

Verified against the real model, with a deliberate contrast

material	medium says	model judged
「第一次办展那天…」 — text, but a story	solve	share ← the broken case
「租房合同这三条…」 — text and a method	solve	solve, and it drafted the two solve fields from the material
model declines	—	null → falls back to the medium

Reverting the change fails the new test (an image item whose caption is an experience must come out share).

Full backend suite: 503 passed.

Not in this PR

The next slice would stop the taxonomy being three values at all: separating what the content is (open, AI-named) from which finite behaviour to run (the routing key). That is a migration — the CHECK constraint, 139 references across 12 files, ADR 0002's NULL semantics, E2E copy — so it wants its own plan rather than being smuggled in here.