[thorwhalen]

Fixes #77.

Behaviour chosen: option 2 (wrap to more lines on a narrow frame)

A still's subject is shared by every cut, but legibility is per-delivery: type is sized as a fraction of frame height and must fit the frame's width, so a 9:16 cut has about half the width per line of a 16:9 one and twice the height to spend. The least surprising fix is the one tituli already implements (shrink, then wrap, never truncate), given the line budget the frame can afford.

• _wrap_labels_for_frame runs just before tituli.video.overlay in the text pass. For each still caption that tituli cannot set in its default line budget, it lays it out with a budget of ceil(3 * (16/9) / aspect) lines, capped at 6 (named constants).
• Only captions that would previously have raised are touched. A caption that fits keeps its payload, so every existing render (landscape or vertical) is byte-identical. No data-model change, no new setting, no schema change, no public API change.
• Raise-never-truncate holds: a caption too long even for the taller budget still raises TextDoesNotFit, still naming the still and the delivery (transforms: name the still and delivery in a caption-overflow error #82). Cards and attributions are untouched.

Why not the others: (1) subject_short is a body-schema addition plus an authoring judgement per still, and is a federation event under this repo's rules. (3) opting out drops the attribution a labelled still exists to show. Both remain available later; this change does not foreclose them.

Measured: the issue's still (a 100-char subject) needs 5 lines at 0.022 on 1080x1920; it now sets in 5 and renders in full (checked on a real 1080x1920 overlay render).

Tests

Three new tests (tall frame wraps what a wide frame refuses; a caption that fits is left identical; too long for even the taller budget still raises) plus doctests. Removing the wrap makes the first and third fail. Full run with --doctest-modules: 709 passed; the 8 failures are all in tests/test_mcp_caller_key.py and tests/test_wire_descriptions.py cannot be collected, solely because py2mcp is not installed in the dev venv (CI installs the mcp extra).

Deploy

Merging releases to PyPI. Production picks it up by: braidio_mcp and reelee_mcp connector redeploys (unversioned git pins) and pip install --no-deps braidio==<ver> in the backend venv, then restarting the backend.

Adversarial review (fresh Opus subagent)

Verdict: OK to merge, no correctness defects. It checked: exact 16:9 sizes keep the default budget (no float creep), a caption that fits returns the identical overlay, no cache-key change needed (key already covers size, delivery and subject), the error path still names the still, attribution overflow still raises, nothing is truncated, size as list or tuple. One confirmed gap: nothing guarded the call site in _composite. Fixed in the second commit with test_a_vertical_finish_hands_tituli_the_wrapped_caption (deleting the call makes it fail).