Fix nested tag fence reassembly vulnerability

Found by the security lane of Task 11635 (unrelated to that task's diff; pre-existing). `apps/api/app/operator_dashboard/review.py` fences each untrusted field of the outcome-review prompt in a tag and strips any copy of the fence tag from the text first (`_FENCE_RE` / `_fenced`, around lines 170-177). The strip is a single `re.sub` pass, so a nested copy survives: a report or task body containing the entity-encoded text that decodes to `</rep</report>ort>` loses only the inner tag and leaves `</report>` in the prompt. Reproduced on develop at 82ea879b0: the fenced output is `a </report> b`. Any runner (which writes the report) or task author (title, body) can therefore close the fence early and plant a fake field, steering the 1-5 outcome score. No data access beyond the score. Fix: strip until a fixpoint (loop `_FENCE_RE.sub` until the text stops changing), or escape `<` inside fenced text instead of stripping tags. Add a test in `apps/api/tests/test_operator_dashboard_quality.py` with the nested-tag input and the entity-encoded form, asserting no fence tag survives. Out of scope: the HTML-to-text pass (`apps/api/app/html_text.py`), the scoring prompt's wording, any other prompt in the API. Done means: - `_fenced` returns text in which no `<tag>`/`</tag>` of the fence name survives for the nested and entity-encoded inputs above; a test pins both. - `./dc check quick --gate-result` passes; `./dc test tests/test_operator_dashboard_quality.py` passes. Required checks: quick gate, test_operator_dashboard_quality.

Outcome-review _fenced now strips fence tags to a fixpoint and also strips unclosed or spaced tags (a security-lane finding). Tests pin the nested and entity-encoded inputs. Quick gate and test_operator_dashboard_quality (30) passed. PR #7453 is queued to the merge lane. Memory: none applied.