fix: check the context of the issue, which injected to the prompt (#38) - #39

FetchIssueContent rendered every comment on the issue into the prompt,
including the daemon's own chatter. The pickup comment is posted before
the fetch, so it landed in the prompt even on the first run; on a re-run
the park comments (with their full error dumps), PR links and stop
notices piled up too, so the model read a transcript of previous runs as
if it were part of the report.

Every status comment now ends with a hidden `<!-- loope:bot -->` marker
and FetchIssueContent drops marked comments. Two judgment calls, made
headlessly:

- needs-info is deliberately NOT marked: it carries the numbered
  questions a human answers in the following comment, so stripping it
  would orphan the answer.
- comments posted before the marker existed are matched by the exact
  opening text of each template (legacyBotStatusPrefixes), so issues
  already carrying chatter — like #38 itself — are cleaned up too,
  without risking a human comment.

The UAT checklist the tool appends to the issue *body* is left in place:
it states what the change must satisfy, so it is context, not noise.

Closes #38