fix: plot_format leaks its settings when the wrapped block raises

[WatchTree-19]

What does this fix?

1. plot_format leaks its settings when the wrapped block raises.

old = get_plot_format()
set_plot_format(plot_format, dpi)
yield
set_plot_format(*old)

There is no try/finally, so if the wrapped code raises, the restore never runs and _PLOT_FORMAT, _PLOT_MIME_TYPE and _PLOT_DPI stay overridden for the rest of the process. Nothing warns, so every subsequent block silently renders at the wrong format and DPI. Against released 1.4.3:

baseline           : ('png', 100) mime = png
caught             : rendering blew up
after failed block : ('svg', 300) mime = svg+xml     <- leaked

2. set_plot_format half-applies an unsupported format.

_PLOT_FORMAT = plot_format
_PLOT_MIME_TYPE = _MIME_TYPES[plot_format]

The assignment happens before the lookup, so a format not in _MIME_TYPES raises KeyError with _PLOT_FORMAT already changed and _PLOT_MIME_TYPE still on the old value:

after set_plot_format("jpeg") : format='jpeg', mime='png'   <- inconsistent

The change

try/finally around the yield, and the mime lookup moved ahead of both assignments so a bad format changes nothing.

Tests

Two added next to the existing test_plot_format_ctx_manager, which only covers the success path. Both fail on master and pass with the change; the full test_image_unit.py module passes (9 passed).

One thing I left alone

set_plot_format with a bad format still raises a bare KeyError: 'jpeg', which is not a great message for what is really an invalid argument. I kept the exception type as-is to avoid changing behaviour anyone might be catching, but happy to turn it into a ValueError naming the supported formats if you would prefer that.

[coveralls]

coverage: 88.607% (-0.2%) from 88.791% — WatchTree-19:fix/plot-format-restore into man-group:master

[WatchTree-19]

CI is red here, but as far as I can tell none of it comes from this change. Flagging what I found in case it is useful.

build_3_8 fails in sudo apt-get update during image setup, before any tests run.

build_3_11 and build_3_12 have 8 failures, all in table formatter tests:

• AttributeError: 'Series' object has no attribute '_append' (6 tests)
• TypeError: Invalid value '' for dtype 'float64' (2 tests)

This PR only touches pybloqs/block/image.py and tests/unit/block/test_image_unit.py.

Both look like pandas 3 turning existing deprecations into errors. On clean master at pandas 2.3.3 those same code paths already emit the matching FutureWarnings, table_formatters.py:801 for the fillna downcasting and the data_rows.loc[:, ~data_rows.columns.isin(columns)] = "" setitem. build_3_9 and build_3_10 pass, which fits, since they resolve an older pandas.

I could not install pandas 3 in my environment to confirm master fails the same way, so treat that last part as inference rather than something I verified. The 9 tests in test_image_unit.py pass on this branch.

Glad to send the pandas 3 compatibility work as a separate PR if that would help.

[WatchTree-19]

Following up on my last comment, I said the pandas 3 part was inference. I have now confirmed it.

I built a Python 3.12 environment with pandas 3.0.5 and ran clean master with no changes:

FAILED tests/unit/block/test_table_formatters.py::test_FmtAppendTotalsRow_modify_dataframe
FAILED tests/unit/block/test_table_formatters.py::test_FmtAppendTotalsRow_mixed_datatypes
FAILED tests/unit/block/test_table_formatters.py::test_FmtAppendTotalsColumn_modify_dataframe
FAILED tests/unit/block/test_table_formatters.py::test_FmtExpandMultiIndex_modify_dataframe
FAILED tests/unit/block/test_table_formatter_builder.py::test_smokey
5 failed, 60 passed, 1 xfailed

So it is pre-existing and unrelated to this PR. I have opened #177 with the fix, which is independent of this one.

build_3_8 is still a separate matter, it dies in sudo apt-get update before any tests run.

[WatchTree-19]

The three red checks here are all pre-existing rather than caused by this change, and two of them are fixed by #177.

This branch touches pybloqs/block/image.py and tests/unit/block/test_image_unit.py only.

build_3_11 and build_3_12 — 8 failures, every one in table-formatter code:

8 failed, 282 passed, 1 skipped, 3 xfailed
tests/unit/block/test_table_formatters.py::test_FmtAppendTotalsRow_modify_dataframe
    AttributeError: 'Series' object has no attribute '_append'
tests/unit/block/test_table_formatters.py::test_FmtAppendTotalsColumn_modify_dataframe
    TypeError: Invalid value '' for dtype 'float64'
tests/unit/block/test_table_formatters.py::test_FmtExpandMultiIndex_modify_dataframe
tests/unit/block/test_table_formatter_builder.py::test_smokey
tests/integration/test_table.py::test_formatters_with_operator_on_df_with_nans_replaces
tests/integration/generate/test_table.py::test_df_to_jinja_table_add_extra_formatters
tests/integration/generate/test_data_tables.py::test_df_to_data_tables_jinja_table_add_extra_formatters

These are pandas 3 removals — Series._append is gone and setitem is strict about dtype. pandas 3 only resolves on 3.11+, which is why build_3_9 and build_3_10 pass and 3.11/3.12 don't.

#177 is the fix for exactly those, and it's green on 3.11 and 3.12. So merging #177 first should clear these; I'll rebase this onto master afterwards.

build_3_8 — the CircleCI image sudo apt-get update failure (Chrome repo, NO_PUBKEY FD533C07C264648F, exit 100). Same step, same error, on #177. Not code-related.

Nothing here needs changing on my side until #177 lands — happy to be told otherwise.