Treat a figure as content, not as a block to ignore

[peterbjohnson]

On UCL_MechEng/Worksheet_2.pdf the spec the agent wrote holds ignore: Para text~'Figure [0-9]+:'. Mathpix writes each figure as one paragraph, the image line then its caption, so that selector marks every figure of the sheet ignored: Figure 1 (the Lucy Ashton, which Q4 is about), Figure 2 (the model test results Q4 ii) asks the student to read) and Figure 3 (the drag-coefficient curve Q4 asks the student to digitise). The set built with no images, and nothing in the run said so.

A figure is content. In in2lambda_agent/spec.py, the spec-writing prompt says that a paragraph holding an image belongs to the question or part it illustrates, and that ignore is for headings, rubric and page furniture, never for a paragraph holding ![. After spec run, the pipeline reads the draft and counts the blocks marked ignore whose lines hold an image reference; if there are any, it reports them as a coverage finding (b10 (lines 29-30) holds an image and is marked ignore) and the spec is rewritten once with that finding, as t16's iteration does for unassigned blocks; if the rewrite still ignores an image, the run carries on and the stage line says how many images were dropped. A test in tests/test_spec.py runs a source with a figure paragraph through a fake backend whose first spec ignores it and asserts that the second call's prompt names the block and that the stage line reports it. Done when a run over Worksheet_2.pdf builds a zip whose media/ holds the three figures, or says which it dropped.

---

Workbench ticket t24.