[ikashapov]

Apply w:pgBorders page borders and the w:docGrid line grid

Closes #151. Based on main (b2c0174, v0.5.1); one commit, branch
issue-151-pgborders-docgrid. Independent of the other open branches.

Two section-level features that change how a page looks, filed together
because both are w:sectPr children; implemented as the issue suggested —
independently — and landing together.

w:pgBorders — §17.6.10 (was: not parsed at all)

Parse + model. New PageBorders on SectionProperties (Dup, like its
siblings), attributes kept optional so the document is mirrored as written;
the renderer applies Word's defaults from [MS-OE376] §2.6.10
(offsetFrom="text", zOrder="front") and §17.6.10's own
(display="allPages"). One parsing subtlety: page borders are the only
place @w:val spans all of §17.18.2 ST_Border — the 27 line styles
plus ~165 art names no other border may carry. The val lands as a string
and classifies afterwards, so val="apples" cannot take down the whole
<w:sectPr>; an art edge is carried losslessly (sz reinterpreted as the
whole points the spec gives art borders, not the line styles' eighths).

Render. A new layout::page_borders pass runs in phase 2 beside
header/footer rendering — the only point where a finished page's command
list and its section's PageConfig coexist, and @w:zOrder needs exactly
that pair: painting order is z-order, so front appends after
everything (headers, body, footers) and back prepends. Geometry per the
ECMA prose and the Word↔Writer interop notes (editeng:: BorderDistanceFromWord): space measures to the border edge nearest
the reference — from the page edge inward, or from the text margin
outward; horizontal edges own the corner squares, vertical edges run
between the bands, the same junction convention the table painter uses.
Edges resolve through the tables' existing convert_model_border, so
double draws its §17.4.38 two-thirds band via the same
(pub(crate)-lifted) emit_border_rect, the other 24 line styles
approximate solid warn-once, and a double page border can never drift
from a double table border.

Declined, and recorded where: art borders draw nothing (the reading
LibreOffice takes) with a warn-once; @shadow/@frame accepted-and-
dropped; w:bordersDoNotSurroundHeader/Footer not read — a text border
always references the configured margins (module doc says why the
configured margin, not the per-page pushed-down content, is the
reference).

w:docGrid — §17.6.5 (was: parsed, then ignored)

The gate that matters most: the pitch applies only when @w:type is
lines/linesAndChars/snapToChars. Every Word document on earth
carries <w:docGrid w:linePitch="360"/> with no type, and §17.18.14
spells out that absent/default means "no document grid" (kept so the
settings survive re-enabling). A test pins this directly, and the 28
existing corpus documents render byte-identically before and after.

The algorithm (LibreOffice's Word-conformed CalcRealHeight, plus
§17.6.5's own overrides): a gridded line consumes ceil(h/pitch) whole
slots — tall text jumps straight to two, Word's signature behavior — with
its natural box vertically centered in them, which is the visible
half-gap above the first line of a Japanese page. lineRule="exact"
escapes (the one override §17.6.5 names); atLeast takes
max(slot, min) except the MSO atLeast-0 quirk (grid off for the
line); auto multiplies the slotted height, clamped below at 1.0 — under
a grid, "1.5 lines" is 1.5 pitches. Registration is by quantized heights
alone (no absolute re-snap), which keeps the advance
position-independent — that is what lets pagination sum precomputed line
heights and still agree with emission to the point.

Where it lives: one function, resolve_line_metrics, feeding the four
sites the module doc already required to agree (fitting, emission, the
border box, the empty-paragraph path — an empty gridded paragraph holds a
slot open, like Word's empty 行). The pitch reaches lines only by being
stamped onto ParagraphStyle in the body page flow, so the spec's scope
comes free: table cells (no w:adjustLineHeightInTable support),
headers/footers, footnotes and shape text never pass the stamp point.
w:pPr/w:snapToGrid (§17.3.1.32) is newly parsed, cascades through the
style hierarchy like its sibling toggles, and opts a paragraph out.
Character pitch (linesAndChars/snapToChars's horizontal half) is not
applied; the module doc records the boundary.

Tests and fixtures

• test-files/page-borders.docx — the same four deliberately distinct
  edges (each its own colour, width, space) through both offset modes,
  one section each; every edge's band is pinned by colour, and the two
  pages differ only in reference frame. test-files/doc-grid.docx —
  identical content twice: type="lines" linePitch="360" vs the
  type-less grid, making section 2 the ungridded control the assertions
  measure against so no font metric is pinned. Both by
  scripts/make_issue151_fixtures.py, verified with verify_docx.py.
• tests/page_borders.rs (8): both fixture pages edge-by-edge, display
  selection, zOrder front/back as command order, art/nil/none
  declines, the text default.
• tests/doc_grid.rs (9): pitch deltas, slot centering as a relation
  between the fixture's two sections, both escapes (exact,
  snapToGrid=0), the type-less no-op, the two-slot jump for 20pt text,
  the empty-paragraph slot, the table-cell exclusion.
• Unit tests: the grid math table (rounding, boundary epsilon, both
  escapes, atLeast/auto interactions, leading-inclusive slot base),
  PageConfig::from_section gating, the pgBorders schema (art vals,
  defaults-as-None, all four edges), the snapToGrid toggle and cascade.
• Mutation-checked: six deliberate breaks (grid off, snap ignored,
  type-less applied, display ignored, centering dropped, keepNext
  measure unstamped) each fail their own tests.
• Corpus: all 28 pre-existing fixtures render byte-identical PDFs
  against main; the two new fixtures verified visually (CoreGraphics
  raster) — page borders in both frames, and the gridded page showing
  the even 18pt pitch against its compact control.

Review

An adversarial multi-agent review of the diff (3 lenses, every finding
independently re-verified against the code) confirmed and led to fixing,
before this PR was opened:

• keepNext × grid (major, found by all three reviewers): the keepNext
  group pre-measure laid paragraphs out with the unstamped style, so a
  gridded chain was measured at natural heights, held at a page bottom it
  could not fit, and torn. The measure now stamps the same pitch the
  placement path does; a regression test builds the exact page geometry
  where the two disagree.
• Endnotes page: it continues the last section's flow (§17.11.2), so it
  now wears that section's borders — at the page index after the
  section's own pages, so display="firstPage" does not re-fire on it
  (both asserted).
• Fixture schema order: CT_SectPr is a strict xsd:sequence; the
  fixtures originally wrote pgBorders/docGrid before pgSz — exactly
  the class of package this parser tolerates and Word refuses (the
  issue-165 lesson). Regenerated in schema order, and the ZIPs are now
  byte-reproducible (pinned timestamps, like the sibling scripts).
• Smaller: the three ST_PageBorder* § citations were off by three
  (§17.18.62–64, cross-checked against the repo's own ST_PageOrientation
  = §17.18.65); two pre-existing wrong § numbers on
  header_refs/footer_refs that collided with the new correct ones were
  dropped per the AGENTS rule; the README coverage tallies were updated;
  a shared-continuous-page border ownership note (with its open
  Word-reference-render question) landed in the module doc.

Gates

cargo fmt --check, clippy --all-targets -- -D warnings,
RUSTDOCFLAGS="-D warnings" cargo doc --no-deps,
cargo build --no-default-features and the full 46-target suite pass.
No new dependencies; public API unchanged.