chore(deps): bump playwright/python from v1.62.0-jammy to v1.63.0-jammy in /tests/NetPdf.RenderingCorpus/docker - #373

…des #373 #376 #377) (#378)

* chore(deps): move to stable AngleSharp.Css 1.1.2 + latest everything

Supersedes Dependabot #373, #376 and #377, and goes further where they lagged:
SkiaSharp 4.153.1 (+ matching HarfBuzzSharp 14.2.1.301), coverlet 10.1.0 and
docfx 2.81.0 are newer than what Dependabot proposed.

The headline is AngleSharp.Css: 1.0.0-beta.144 -> 1.1.2, the first STABLE
release NetPdf can take. 1.0.0 and 1.0.1 were held because they broke
background-image painting. 1.1.2 does not: all background tests and every
rendering golden (LayoutSnapshots, PaginationGolden, RenderingCorpus) are
byte-identical. Every runtime dependency is now stable, so the NU5104
prerelease-dependency exception is removed and that check is an error again.

Both Dependabot PRs failed all five enforcing legs. Each failure, run down:

1. REAL REGRESSION, fixed. W3C Flexbox + Grid gates failed
   (flex-gap-main-axis, flex-gap-with-justify-end, grid-gap-shorthand-columns).
   AngleSharp.Css 1.1.x expands a single-value `gap: 20px` into
   `row-gap: 20px` + an EMPTY `column-gap` instead of repeating the value
   (CSS Box Alignment L3 8.3). NetPdf drops empty longhands, so every
   one-value flex row and grid silently lost its column gap. CssParserAdapter
   now repairs an empty column-gap / grid-column-gap from its row twin,
   carrying the importance. Safe: an empty standard longhand is never
   authored CSS, and a `gap: var(--g)` expansion leaves row-gap empty too, so
   the var() recovery path is unaffected. New tests pin 1-value, !important,
   grid-gap and the untouched 2-value form.

2. BEHAVIOUR GAIN, tests updated. 1.1.x parses @container and @layer as real
   grouping rules (the beta dropped them; only the preprocessor's opaque
   RawBody survived). The cascade already handles both shapes:
   - @container bodies are still SKIPPED with
     CSS-CONTAINER-QUERY-UNSUPPORTED-001. A new CascadeCorpusTests case pins
     this end to end, because with real child rules a naive grouping-rule
     walk would have started applying container styles unconditionally.
   - @layer bodies now APPLY in layer order, where they used to be dropped.
     The corpus test now asserts the unlayered rule still beats the layered
     one (Cascade L5 6.4) and that CSS-AT-RULE-UNKNOWN-001 no longer fires.
   The adapter tests now assert "the body survives in some form" rather than
   the beta's exact opaque shape.

3. KNOWN GAP CLOSED upstream. GridShorthandProductionTests pinned
   `grid-row-end: 6 !important; grid-row: 2 / 4` at end=4 because the beta's
   per-rule dedup discarded the !important longhand. 1.1.x keeps it; the test
   is flipped to the spec-correct 6, exactly as its own comment instructed.

The Directory.Packages.props note now records all three facts so the next
bump starts from them.

Verified locally: build 0 errors / 0 warnings; UnitTests 8618 passed /
3 skipped; RealDocuments 106; W3cConformance 6; RenderingCorpus 41;
LayoutSnapshots 30; PaginationGolden 1; PdfValidation 1; fuzz smoke 0
findings; AOT/JIT parity verified; pack clean with no NU5104.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(css): exact gap expansion from raw CSS; recover @layer bodies

PR #378 review follow-up. All 7 threads were valid.

[P2 x2] The single-value gap repair picked ANY surviving row-gap as the
column value. For `gap: 20px; row-gap: 30px` it produced column-gap: 30px;
the correct value is 20px. Probing AngleSharp.Css 1.1.2 showed why no
adapter-side fix can be exact: for that rule it keeps only row-gap: 30px and
an empty column-gap — the shorthand's 20px is gone from its object model.

So `gap` / `grid-gap` are now re-expanded by the preprocessor from the RAW
declaration (new GapShorthandExpander, paren-aware so calc()/var() stay one
component) as shorthand-expansion recovery records. The adapter's existing
source-order + !important merge then decides each longhand exactly like the
cascade: pinned for gap-then-row, row-then-gap, important shorthand, gap-then
-column, column-then-gap, and calc(). grid-gap emits the modern row-gap /
column-gap names, the only ones the engine registers. The adapter repair
stays only as a fallback for paths with no recovery data, and now carries
the empty entry's own importance (the shorthand's), not the row twin's.

[P2] @layer children skipped the recovery pass. The cascade now APPLIES
layer bodies, but CssPreprocessor.GroupingAtRules excluded `layer`, so the
body was captured as an opaque blob and child rules fell back to
AngleSharp-only adaptation — modern colors, var()-bearing shorthands and gap
inside a layer were lost. `layer` is now a grouping rule; a test proves the
gap case inside a layer comes out right. @container stays opaque on purpose
(the cascade skips its body).

[Copilot] The reference-generator Dockerfile pip-installed playwright==1.49
inside the 1.63 image, so chromium.launch() would look for a browser the
image does not contain. Pinned to 1.63.0, with a note that the two move
together.

[Copilot x3] Stale text: the PDFtoImage/SkiaSharp note (rewritten without
hard-coded patch numbers so it cannot drift again), and two test comments
that still described the old AngleSharp behavior.

Verified: build 0 errors; UnitTests 8625 passed / 3 skipped; RealDocuments
106; W3cConformance 6; RenderingCorpus 41; LayoutSnapshots 30;
PaginationGolden 1; PdfValidation 1; fuzz smoke 0 findings.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>