Read PPTX merged-cell attributes from a:tc, not a:tcPr

[Yi-111-a]

Why

#1074. In DrawingML, gridSpan, hMerge and vMerge are attributes of a:tc
(CT_TableCell); a:tcPr (CT_TableCellProperties) has no such attributes.
pptx_xml_to_text read them only from a:tcPr, so a merged table written by any
real producer was extracted as if every cell were plain:

| Summary | COVERED | Total |      <- actual
|  | Summary | Total |          <- expected

What changes

• pptx_xml_to_text reads the three merge attributes off a:tc in the Start
  handler, and the two a:tcPr handlers are dropped — they existed only for that
  lookup. The shared bit is a small apply_merge_attrs closure, so the Start
  path and the old duplicated blocks no longer drift.
• The merge fixtures in pptx_tables.rs put the attributes on a:tc instead of
  a:tcPr. They previously exercised the same wrong structure the parser
  expected, which is why the existing tests could not catch this.
• New merged_cell_attributes_on_a_tc_preserve_the_grid_shape spells the markup
  out directly rather than going through the helpers, so it keeps testing a real
  producer's shape (<a:tc gridSpan="2">) if the helpers change later.
• a_vertical_merge_continuation_stays_an_empty_cell used a textless
  continuation cell, so clearing its text was a no-op and it passed whether or
  not the vMerge lookup worked. That cell now carries the text PowerPoint
  repeats, and the assertion pins that it is dropped.

No a:tcPr fallback is kept: those are not a:tcPr attributes in any version of
the schema, so reading them there was the bug. Say the word if you would rather
tolerate producers that write them there anyway — it is three lines.

How it was checked

cargo +1.95 test -p utopia-ingest --lib --test pptx_tables --test pptx_order --test pptx_breaks --test office_text_references → 108 + 12 + 4 + 2 + 4 passed,
0 failed. The three merge tests are the discriminating ones: reverting only
parsers.rs fails exactly those three and passes the other nine, and the failure
output is the | Summary | COVERED | Total | line quoted above. cargo fmt -p utopia-ingest -- --check is clean.

Two things I did not run, so CI is the gate on them:

• cargo clippy --workspace --all-targets -- -D warnings reports two
  collapsible_match findings in parsers.rs, and both reproduce on unmodified
  dev (parsers.rs:313 and the a:t arm at parsers.rs:1107 there, 1110
  with this change) — neither is in code this PR touches, and it adds no new
  findings. I left them alone rather than mixing an unrelated lint cleanup into a
  bugfix; note CI pins stable, so if your stable predates these they may not
  fire for you.
• cargo test --workspace. This change is confined to pptx_xml_to_text, so I ran
  the ingest tests that cover it, but the full workspace build did not fit the
  disk I had (it filled the volume and the linker failed on space). Worth a
  look on CI, particularly the utopia-store suites that need
  UTOPIA_DATABASE_URL.

Before review

• Every commit is signed off (git commit -s)
• cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings and cargo test --workspace pass
  (fmt and the ingest tests do; clippy and the full workspace run do not — both
  explained above, neither caused by this change)
• For changes under web/: pnpm build and pnpm test pass — not applicable, no web/ change
• SQL under crates/utopia-store/ was tested with UTOPIA_DATABASE_URL set — not applicable, no SQL change
• A new migration takes the next free number on dev — not applicable, no migration
• UI strings are in both web/src/i18n/en.ts and zh.ts — not applicable, no UI string
• A change to the data model, the ontology contract or a public API has its ADR in docs/decisions/ — not applicable

Disclosure

Written with opencode (space-bunny-free). The parser change, the fixture
changes and the tests are all generated; the diagnosis, the DrawingML
CT_TableCell reference, the choice to drop the a:tcPr path and the decision to
strengthen the vMerge test are the issue reporter's and mine. It needs a human
read before merge — in particular please sanity-check that dropping the a:tcPr
lookup is what you want rather than a tolerated fallback.

[WaylandYang]

Thanks @Yi-111-a. This is the same fix as #1077, which @Floating-Y opened three hours earlier together with the issue, so I am landing that one and closing this. Your reading of CT_TableCell against CT_TableCellProperties is right, and it is the reason the fix is correct.

If you would like something that is not taken: the fixtures in pptx_tables.rs are hand-written XML, which is how this slipped through. A test that reads a deck written by a real producer would have caught it, and the same goes for the Word and spreadsheet readers.