feat(tables): merge robustness — Table.merge_cells/split_cells + Cell inspection accessors (Tables 2.0 Phase 3)

[MHoroszowski]

Phase 3 of issue #12 (Tables 2.0): merge robustness API

Closes the third sub-feature on the Tables 2.0 epic. Phase 1 (PR #37) shipped row/column add/remove. Phase 2 (PR #40) shipped the table style API. This PR adds the range-style block-merge and inspection surface, leaving sizing/ergonomics for a follow-up Phase 4.

Before this change the only merge surface was Cell.merge(other_cell) — two-cell, raises ValueError on any range overlap, no idempotency. Power users hit this every time they programmatically built tables with block merges; they ended up wrapping every merge in try/except or pre-checking with is_merge_origin/is_spanned. Inspection of the underlying gridSpan / rowSpan / hMerge / vMerge was only reachable via the private _tc element.

Public surface added (additive — existing Cell.merge and Cell.split keep current behavior)

# Read-only inspection on _Cell (snake_case per module convention)
cell.grid_span            # 1 by default; >1 only on horizontal merge origin
cell.row_span             # 1 by default; >1 only on vertical merge origin
cell.h_merge              # True only on cells horizontally spanned
cell.v_merge              # True only on cells vertically spanned

# Range-style block merge
table.merge_cells((0, 1), (0, 2))           # tuples = inclusive (rows 0..1, cols 0..2)
table.merge_cells(range(0, 2), range(0, 3)) # range objects = half-open
table.merge_cells((0, 1), (0, 2))           # idempotent — second call is a no-op

# Range-style split (inverse)
table.split_cells((0, 1), (0, 2))           # un-merge a contained block
table.split_cells((0, 0), (0, 1))           # raises ValueError if merge extends beyond range

What it adds

_Cell read-only inspection accessors

• Cell.grid_span / row_span / h_merge / v_merge — read-throughs to tc.gridSpan / tc.rowSpan / tc.hMerge / tc.vMerge. Snake_case per the existing _Cell convention (is_merge_origin, is_spanned, span_height, span_width, margin_top, etc.). The issue body used the OOXML attribute names literally (camelCase) but those are XML serialization details, not the Python public surface. Forge audit (pre-commit) flagged the original camelCase as a permanent API contract violation; renamed before merge.

Table range-style block merge

• Table.merge_cells(row_range, col_range) — row_range and col_range accept either a 2-tuple (start, end) (inclusive — (0, 1) covers rows 0 and 1) or a Python range object (half-open per Python convention). Order within either form is irrelevant. Idempotent on exact re-merge: a second call with the same range on an already-merged-exactly-this-way region is a no-op and returns the existing merge-origin _Cell. Single-cell range is also a no-op. Returns the merge-origin _Cell. Raises ValueError on partial overlap with a different-shape existing merge.
• Table.split_cells(row_range, col_range) — inverse operation. Idempotent on un-merged ranges. Splits any merges fully contained in the range. Raises ValueError if a merge in the range extends beyond the boundary (would orphan the rest).

Helpers (private)

• _normalize_range(rng) — single normalization point: tuple → inclusive (low, high), range → inclusive low/high (high = stop-1). Raises TypeError on other shapes; raises ValueError on negatives, empty range, or non-unit step.
• _find_merge_origin(tbl, r, c) — two-pass walk (left through hMerge, then up through vMerge) used by split-boundary validation. Verified by Forge to land correctly on the origin for h-only, v-only, and block-inner spanned cells.

Out of scope (deferred)

• Replacing or deprecating Cell.merge(other_cell) / Cell.split() — they stay as-is.
• Cross-table merges (both ranges must be in the same Table instance).
• Diagonal or non-rectangular merges (PowerPoint's data model is rectangular only).
• Per-merge formatting carryover beyond what Cell.merge already does (move_content_to_origin continues to handle text-content collation; we don't add new merging strategies).
• Phase 4: per-row height / per-column width round-trip + Table.row_count / column_count ergonomics.

Tests

• 46 pytest cases in tests/test_tables_phase3.py covering each new accessor, every public path through merge_cells / split_cells (including partial-overlap, boundary-cross, single-cell, range-object, order-agnostic, multiple-merges-in-one-call), the _normalize_range helper (parametrized), round-trip preservation through save/reload, and Phase-1/Phase-2 regression checks. Full suite: 3404 passed.
• 8 behave scenarios in features/tbl-merge.feature exercising block merge, idempotent re-merge, range-object form, partial overlap ValueError, single-cell no-op, block split, idempotent split, and boundary-cross split ValueError. Full behave: 1029 scenarios passed, 0 failed (baseline 1021 + 8 new).
• Ruff clean: ruff check src tests → All checks passed; ruff format --check → no diff.

Reporting contract (CLAUDE.md §7)

$ python3 -m pytest tests/ -q | tail -3
........................................................................ [ 99%]
....................                                                     [100%]
3404 passed in 4.66s

$ python3 -m ruff check src tests | tail -3
All checks passed!

$ python3 -m behave features/ --no-color 2>&1 | tail -3
1029 scenarios passed, 0 failed, 0 skipped
3100 steps passed, 0 failed, 0 skipped
Took 0min 1.621s

UAT

• uat_tables_phase3.py (untracked per CLAUDE.md §6) at repo root.
• Generates uat_tables_phase3_out.pptx with three tables: a 2×3 block merge, an idempotent merge+split round-trip, and a mixed-shape merge with header bar / vertical merge / bottom-right block.
• All round-trip assertions pass programmatically.
• Visual UAT in PowerPoint signed off by maintainer.

Refs #12