Edit table cells and rows, picture crop, z-order and run hyperlinks from rpptx and Python

[hadim]

Summary

One commit per checklist item of #169, then the archive re-record.

• Table cells (F), binding only. Python Cell gains python-pptx's
  merge(other_cell), split(), is_merge_origin, is_spanned,
  span_height and span_width over the facade's staged merge_to and
  split, plus fill (a FillFormat through a new table cell fill target)
  and margin_left, margin_right, margin_top and margin_bottom. Merge
  and split keep the rectangular grid, so no handle goes stale and the
  revision does not advance.
• Row heights and cell borders (H1). TableRef and TableMut read
  row_height, and TableMut::set_row_height stages the change like
  set_column_width and keeps the graphic frame height equal to the sum of the
  rows. Both setters now share one checked extent helper. The new
  CellBorder enum names an edge, TableCellRef::border and
  TableCellMut::border read it, and TableCellMut::set_border replaces or
  clears it. Python gains Table.rows (a RowCollection of Row handles with
  a writable height) and Cell.border_left, border_right, border_top
  and border_bottom as LineFormat views, reusing the width, colour and
  fill logic of shape lines.
• Picture crop (J). ShapeRef::crop and ShapeMut::set_crop read and
  write a picture's a:srcRect as four Percent1000 insets, now re-exported
  from rpptx. Python Shape gains python-pptx's crop_left, crop_top,
  crop_right and crop_bottom floats. No oxml-drawing change, so no
  overlap with Paint slide backgrounds in PDF output and open gradients without ang or path #175's fill.rs work.
• Z-order (I). CT_ShapeTree::move_child(from, to) in rpptx-oxml moves a
  child's bytes past exactly the children between its two indices and
  reparses, as remove_child_by_id does, so unmodelled members such as
  p:contentPart and a schema-final p:extLst keep their bytes and place.
  Both now share one range finder. Presentation::move_shape(slide_index, from_index, to_index) exposes it with move_slide's final-index semantics,
  and Python gains ShapeCollection.move(from_, to).
• Run hyperlinks (L1). Presentation::hyperlink_address(slide_index, relationship_id) resolves the r:id of an a:hlinkClick, and
  Presentation::set_run_hyperlink(slide_index, shape_id, paragraph_index, run_index, address) adds, retargets or removes a run's click hyperlink in
  one staged step, reusing the slide's relationship to the same address and
  pruning the old one once nothing on the slide names it. Python gains
  Run.hyperlink, a Hyperlink view whose address follows python-pptx.
• Archive rows. rpptx and rpptx-oxml re-recorded in
  scripts/readme_doctests.py and their READMEs.

Every new Python name is in _rpptx.pyi and typing_smoke.py. HLD 06
describes the new facade methods and HLD 10 the new binding surface. The
rpptx-py README capability list gains one line per feature.

Part of #169: checklist items "Table cells: merge_to / split, set_fill,
set_margins", "Table rows and columns: row height, cell borders", "Z-order",
"Picture crop" and the run half of "Hyperlinks on runs and shapes". Still open
after this PR and #181 (items A to E):

• adding and removing table rows and columns (H2, the largest item, needs an
  oxml-drawing table change),
• inherited placeholder geometry (G, which the Production readiness for editing real docx and pptx files: an acceptance contract, two realistic fixtures and two matrices #158 "shape moved and resized"
  row still hits),
• changing a slide's layout (K),
• shape click hyperlinks (L2), which can reuse hyperlink_address and the
  relationship reuse and pruning of set_run_hyperlink,
• comments anchored on a shape or text range (M) and built-in table styles (N),
  proposed as out of scope for this round.

Why

These are the #169 items the facade either had and Python did not bind (F),
or that the model parsed but no public API reached (H1, J, L1), or that did
not exist at all (I). A deck pass that merged a header, sized a row, drew a
cell edge, cropped a picture, brought a shape forward or linked a run had to
keep a python-pptx step for it.

Notes

• API choices: cell margins read None when absent,
  borders are per-edge LineFormat views, z-order is shapes.move(from_, to)
  with one revision bump as a structural change, and an unreferenced hyperlink
  relationship, and any relationship only the old hyperlink named such as an
  a:snd click sound, is pruned on retarget or removal.
• Maintainer decision, recommended answer "keep the atomic call": an alternative
  is relationship-level primitives (relate_hyperlink, then write the
  run, then prune). While testing them I found that validate() reports an
  unreferenced hyperlink relationship as DanglingRelationship, and
  to_bytes debug-asserts a clean validation. A public relate-then-write API
  would therefore leave a window where the deck does not validate, and a
  caller who related an address without using it would keep an invalid deck.
  set_run_hyperlink does the relationship and run changes together on a
  staged slide clone and relationship list, and a rejected call changes
  nothing. It addresses the run by shape id, like picture_image and
  replace_picture_image, and refuses an id that another slide child shares,
  so a deck with duplicate ids (which validate() already reports) raises
  rather than editing the wrong shape. hyperlink_address stays a read-only
  primitive that L2 can reuse for shape click hyperlinks.
• Maintainer decision, recommended answer "keep": a native refusal raises
  RpptxError, not the ValueError python-pptx raises, for an overlapping
  merge, a split of a cell that is not a merge origin, a row height that is not
  positive, and a hyperlink address with a control character. This matches the
  existing Column.width setter. Binding-side checks (a cell of another table,
  a crop value out of range, a margin outside 32 bits, a non-picture crop)
  raise ValueError.
• Maintainer decision, recommended answer "keep": Row.height also updates
  the graphic frame height, as set_column_width already updates its width.
  python-pptx changes only a:tr/@h.
• Behaviour details worth a look: set_crop keeps the stored attribute of an
  unchanged edge, drops a changed edge of zero, and adds no a:srcRect for
  four zero insets on a picture without one (python-pptx would write an empty
  a:srcRect). Assigning a run's current hyperlink address changes nothing and
  keeps any tooltip, and any other assignment writes a fresh a:hlinkClick
  with only r:id, as python-pptx does. Relationships the old element named
  and nothing else does, its click sound included, are removed. Reading a cell border or a cell fill
  never creates a:tcPr children, and clearing an absent margin leaves the
  bytes unchanged.
• API changes, all additive, no breaking change: rpptx gains CellBorder,
  the Percent1000 re-export, TableRef::row_height,
  TableMut::row_height/set_row_height, TableCellRef::border,
  TableCellMut::border/set_border, ShapeRef::crop, ShapeMut::set_crop,
  Presentation::move_shape, hyperlink_address and set_run_hyperlink.
  rpptx-oxml gains CT_ShapeTree::move_child. Two behaviour-preserving
  refactors: set_column_width uses the shared extent helper (same checks and
  messages, covered by the existing table tests) and remove_child_by_id uses
  the shared range finder (covered by the removal tests).
• Not included: Row.cells (python-pptx has it, rpptx Python bindings: what is left for a production deck chain, as one checklist #169 did not ask for it),
  hyperlinks on table cell runs (the Python text surface covers shape text
  frames only) and a:hlinkMouseOver.
• Merge order: the Python tests form their own groups next to related tests
  (table cells and rows after the fill and line test, crop and z-order after
  the picture test, hyperlinks after the text property oracle test), away from
  the insertion points of Keep rpptx handles valid when Python sets a run's text #173 and Bind the missing rpptx facade methods of the deck workflow in Python #181. git merge-tree against Keep rpptx handles valid when Python sets a run's text #173, End the CLIs cleanly on a closed pipe and refuse silent overwrites #174,
  Paint slide backgrounds in PDF output and open gradients without ang or path #175, Expose comparison options as keywords of Python Document.compare #176, Make the body read walkers reach content controls #177, Save documents and presentations atomically, keeping links and modes #178, Align split_run, bookmarks and story comments on the direct body index #179, Keep every child of a text box that replacement rewrites #180 and Bind the missing rpptx facade methods of the deck workflow in Python #181 is clean on the feature commits.
  The archive commit touches the rpptx row that Paint slide backgrounds in PDF output and open gradients without ang or path #175 and Save documents and presentations atomically, keeping links and modes #178 also
  re-record, so whichever lands second re-runs the measurement.

Tests

Added to crates/rpptx/tests/integration.rs:

• row_heights_and_cell_borders_round_trip_with_the_frame_height_in_step:
  row height and frame height after set_row_height, rejected heights leave
  the table unchanged, borders set, cleared and reopened, and lnL, lnB
  before the cell fill in the XML.
• picture_crop_round_trips_and_rewrites_only_the_changed_edges: four zero
  insets add nothing, a non-picture is refused, the reopened crop, a:srcRect
  between the blip and the fill mode, and the edge rules on a hand-written
  a:srcRect.
• moving_a_shape_changes_the_draw_order_and_keeps_every_shape_id: the pixel
  under two overlapping rectangles changes colour in the rendered PNG, ids
  keep their values, bad indices are refused, and the reopened deck validates.
• run_hyperlinks_reuse_resolve_and_prune_their_relationships: rejected
  addresses and runs change nothing, one relationship per address, retargeting
  prunes, removal prunes only when unused, the layout relationship is never
  pruned, a run inside a group, and a shared shape id is refused.
• Four #[ignore] pinned python-pptx 1.0.2 oracles in the style of
  add_table_matches_pinned_python_pptx_table_semantics, through one new
  helper python_pptx_1_0_2_reads: row heights and borders, crop, z-order
  names and run hyperlink addresses as python-pptx reads them.

Added to rpptx-oxml's existing shape_tree unit test module:
moving_a_child_passes_only_the_children_between_its_indices, with a raw
member between children and a schema-final p:extLst.

Added to crates/rpptx-py/tests/test_documented_examples.py, each checked
against python-pptx 1.0.2 in both directions:

• test_table_cells_merge_split_fill_and_margins_like_python_pptx
• test_table_row_heights_and_cell_borders_write_what_python_pptx_reads
• test_picture_crop_matches_python_pptx_in_both_directions (a two-colour
  picture renders differently once cropped)
• test_shapes_move_changes_the_z_order_and_stales_handles_once
• test_run_hyperlink_address_reads_writes_and_prunes_like_python_pptx
  (repeated retargeting keeps one relationship per live address, and the PDF
  carries the new URI)

typing_smoke.py gains one small function per feature.

Run on this branch (macOS arm64, Python 3.12.14, python-pptx 1.0.2):

• cargo fmt --all --check: pass.
• cargo clippy -p rpptx-oxml -p rpptx -p rpptx-py --all-targets --all-features -- -D warnings: pass, also on each commit alone.
• cargo test -p rpptx: 84 lib and 238 integration tests pass.
  cargo test -p rpptx --test integration -- --include-ignored on the eight
  new tests, the four python-pptx oracles included: pass.
• cargo test -p rpptx-oxml: pass (16 unit, 155 integration).
• cargo test -p rpptx-py with DYLD_LIBRARY_PATH set to the libpython the
  test binary links: pass, 0 Rust tests.
• Consumers: cargo test -p rpptx-cli passes except
  validate_rejects_corruption_and_accepts_the_pinned_corpus (missing
  /corpus/pptx, environment only). cargo check --target wasm32-unknown-unknown -p rpptx-wasm: pass.
• maturin develop --locked then pytest crates/rpptx-py/tests: 45 passed, 0
  skipped. Each commit alone: 41, 42, 43, 44 and 45 passed.
• mypy --strict crates/rpptx-py/tests/typing_smoke.py crates/rpptx-py/python/rpptx: no issues. python -m mypy.stubtest rpptx:
  no issues.
• python3 scripts/hash_harness.py --check: 49 entries match.
• python3 scripts/prose_check.py: 0 violations.
  python3 scripts/readme_doctests.py: pass after the re-record.
  python3 scripts/sync_agent_skills.py --check: in sync.

workflow_pptx.py from #158 on fixture-deck.pptx gives the same rows as
main: the geometry, duplicate, replacement and comment rows belong to G and
to #181, and validate needs the CLI on PATH. A probe on the same fixture
merges two cells of the slide 4 table, sets a row height, a bordered, filled
and inset cell, crops the slide 1 picture, moves it to the back and links a
run on slide 2 twice. The result reopens in rpptx, rpptx validate passes,
the PDF carries the second URI, and python-pptx reads the span, the row
height, the margin, the crop, the new order and the link.

Environment-only failures seen: the missing-corpus test above, and the known
libpython load path of the rpptx-py Rust test binary without
DYLD_LIBRARY_PATH.