Adds the regression-gate workflow + actual stable benchmarks to back
it. Previously the benchmarks/ project only contained the one-shot
pre-impl spike harnesses; the v1.0 DoD also requires CI-runnable
benchmarks with a regression threshold.

New benchmarks (benchmarks/NetXlsx.Benchmarks/Benchmarks.cs):
- WriteBenchmarks.ColdCreateAndSave — OPC packaging cost.
  Design §5 target: < 50ms.
- WriteBenchmarks.Write5kRows — in-memory mixed-type write
  (proxy for the 30k-rows-in-memory §5 target).
- WriteBenchmarks.StyledWrite_SmallPalette — CellStylePool dedup
  path (decision #4 / spike 1). 10k cells × 5 distinct styles.
  Design §5 target: > 500k styled cells/s.
- WriteBenchmarks.StreamingWrite_50kRows — SXSSF streaming path
  (proxy for the 1M-rows-streaming §5 target).
- ReadBenchmarks.OpenAndReadColumnSum — open + read 1k rows.

All use ShortRun config (3 warmup × 3 iterations) so the full suite
completes in ~45s on a modern laptop.

New tooling (benchmarks/compare-bench.py):
- Reads BDN's brief-JSON output.
- Compares per-benchmark Statistics.Mean against a baseline.
- Exits 1 if any benchmark regresses > threshold (default 15%,
  configurable via --threshold).
- Reports improvements informationally but never fails on them.
- Tolerates missing baseline (first run) and reports new/removed
  benchmarks.

New workflow (.github/workflows/bench.yml):
- Triggers on PRs and pushes touching src/, benchmarks/, or
  Directory.{Build,Packages}.props.
- Caches a CI-hardware baseline keyed by hash of the source tree.
  Main pushes refresh the cache; PRs read it.
- Runs the comparator with a 15% threshold (10% design DoD + 5%
  CI noise headroom).
- Uploads results + baseline as artifacts for every run.

Two-baseline model documented in benchmarks/README.md:
- benchmarks/baseline/ — committed; dev-local reference for the
  "did my change regress on my machine" pre-PR check. Captured on
  a 12th-gen i9-12900 (fast hw, intentionally not the CI baseline).
- CI cache — managed by the bench workflow; tied to GH Actions
  ubuntu-latest hw. The actual regression gate. Refreshed on every
  main push that touches a tracked path.

The threshold sits at 15% (vs design's 10% DoD) because
short-run BDN measurements on CI runners have higher variance than
full statistical runs. Tighten back to 10% once we have a few
months of CI-baseline data to calibrate against.

Validation: local run produces 5 benchmarks (4 write + 1 read) in
~45s; comparator against the just-stored baseline correctly reports
0 regressions, 0 improvements, 5 unchanged.

Adds multi-run formatted-string support to random-access cells. New
public value records mirror the CellStyle conventions (immutable,
structural equality, nullable-axis "inherit existing" semantics):

  - RichText            — sequence of one or more RichTextRun
  - RichTextRun         — (Text, RichTextStyle) pair
  - RichTextStyle       — font-only subset of CellStyle

Per-run style is restricted to font axes (Bold, Italic, Underline,
FontName, FontSize, Color). Excel's OOXML <r><rPr> element does not
honor per-run fills, borders, alignment, or number format — exposing
the full CellStyle on a run would silently drop those axes. Cell-level
visual style continues through ICell.Style(CellStyle).

Run fonts are pooled through the existing CellStylePool font cache
(decision #4) — runs with identical font properties share one IFont
across the workbook.

GetRichText() returns non-null only when the cell carries explicit
formatting runs (NumFormattingRuns > 0). A cell set via SetString
returns null even though NPOI stores every string cell as an
XSSFRichTextString internally.

IStreamingCell.SetRichText is intentionally absent. NPOI's SXSSF
SheetDataWriter (NPOI 2.7.x) reconstructs a fresh XSSFRichTextString
from cell.StringCellValue at flush time, dropping all in-memory
formatting runs. Per decision #7 (streaming type-honesty), the
absence of the method mirrors the absence of the capability rather
than silently degrading or throwing on call. Surfaced during impl by
a streaming round-trip test; documented in implementation-notes.md.

Surface (30 entries added to PublicAPI.Unshipped.txt):
  - NetXlsx.RichText (+ Runs, PlainText, ctors, structural equality)
  - NetXlsx.RichTextRun (+ Text, Style)
  - NetXlsx.RichTextStyle (+ Bold, Italic, Underline, FontName,
                            FontSize, Color, Default)
  - ICell.SetRichText(RichText)
  - ICell.GetRichText() -> RichText?

Tests (14 new in tests/NetXlsx.Tests/RichTextApiTests.cs):
  - value-type semantics (defaults, equality, validation)
  - in-memory round-trip (SetRichText → GetRichText)
  - file round-trip via Workbook.Open
  - GetRichText returns null for plain strings + non-string cells
  - MaxCellTextLength enforcement on concatenated plain text
  - Zero-length runs contribute no formatting
  - IStreamingCell type-honesty: reflection asserts no SetRichText

Public-API snapshot + DisposedWorkbookMatrix updated.

Design row I-50 added in docs/design.md §6.8.1 with rationale and the
SXSSF streaming caveat. Implementation-notes entry captures the SXSSF
serializer limitation, the font-only design choice, and the read-side
"rich vs plain" discriminator.

First slice of the v1.1 roadmap (rich text → tables → images →
protection → validation → autofilter → named styles → custom
converters → strict concurrency).

Adds a workbook-scoped name registry for CellStyle values.
Register once with a name; apply by name on cells or ranges.

Public surface (~6 PublicAPI.Unshipped entries):
  - IWorkbook.RegisterStyle(name, CellStyle)
  - IWorkbook.GetRegisteredStyle(name)  // null if unknown
  - IWorkbook.RegisteredStyleNames
  - ICell.ApplyNamedStyle(name)
  - IRange.ApplyNamedStyle(name)

Names are case-insensitive. Re-registering an existing name
replaces the definition. ApplyNamedStyle throws ArgumentException
with a friendly message ("No style is registered under 'X'. Use
IWorkbook.RegisterStyle before referencing the name.") when the
name is unknown — distinct from passing a null reference.

Backed by the existing style-pool dedup (decision #4) — equal
CellStyle values still share one underlying NPOI ICellStyle, so
naming is purely a caller-side convenience.

v1.1 named styles are an in-process convenience, NOT OOXML
named-style table entries. When a workbook is saved and reopened
via Workbook.Open, the per-cell visual style is preserved (via
style-pool dedup), but the name -> style map is not rehydrated —
Excel itself never saw the names. Real OOXML named-style table
integration (so styles persist by name across open/save) is
deferred to v1.2.

Coverage (+13 unit tests in tests/NetXlsx.Tests/NamedStyleTests.cs):
  - Empty registry by default
  - Register + get round-trip
  - Case-insensitive lookup ("Header", "header", "HEADER" all match)
  - GetRegisteredStyle returns null for unknown name
  - RegisterStyle replaces existing
  - RegisteredStyleNames enumerates all
  - Reject null name, empty name, null style
  - ICell.ApplyNamedStyle sets the cell's style
  - ICell.ApplyNamedStyle throws ArgumentException for unknown name
  - IRange.ApplyNamedStyle sets every cell in the range
  - IRange.ApplyNamedStyle throws ArgumentException for unknown name

Disposed-workbook matrix updated for the three IWorkbook members
+ the ICell + IRange ApplyNamedStyle methods.

Decision I-57 added to docs/design.md §6.2.2.

Closes the v1.0 external-review recommendation #4 ("commit a date
for NPOI 3.x evaluation or kick off v2 OOXML R&D in earnest").

The date commitment already exists — Spike 4-Q (NPOI AOT/trim
re-check) and Spike 5-Q (NPOI OSMF posture re-check) in
docs/scheduled-spikes.md are due 2026-08-16. What was missing
was a v1.1-completion checkpoint confirming the cadence holds
and recording the additional signal v1.1 produced.

Added entries:

docs/scheduled-spikes.md:
  - Spike 4-Q table: 2026-05-22 row (no re-test, cadence holds).
  - Spike 5-Q table: 2026-05-22 row noting the three NPOI-2.7.3
    surprises v1.1 had to work around (XSSFTable.CreateColumn
    broken on fresh tables, ProtectSheet(null) means unprotect,
    XSSFSheet.RemoveTable absent). Verdict: hold pin; cadence
    holds; next re-assess 2026-08-16.

docs/v2-ooxml-planning.md:
  - Status section gains a 2026-05-22 checkpoint summarizing the
    v1.1 NPOI-surprise count as additional signal in favor of
    evaluating alternatives when the upstream gates fire.
  - Next-milestone language tightens: the August spike re-checks
    are the explicit gating event for whether R&D-1 parallel
    spikes start.

No code changes. No design row needed (these are operational
schedule artifacts, not surface decisions).

Model xl/styles.xml on the SDK engine via a new OoxmlStylePool that dedups
CellStyle values to a single cellXfs index (decision #4) and emits OOXML
schema types (font/fill/border/numFmt/xf) directly. Wires the styling
surface end to end:

- ICell.Style (merge: non-null axes overlay, null inherit), NumberFormat,
  GetStyle, ApplyNamedStyle; IRange.Apply/ApplyNamedStyle (dense);
  IColumn.Width/WidthUnits/Hidden/SetDefaultStyle/ForEachPopulated;
  IRow.HeightInPoints/Hidden; IWorkbook style-pool diagnostics + in-memory
  named-style registry.
- Unblocks the deferred date/time setters: SetDate (DateTime/DateOnly) /
  SetTime / SetDuration write the Excel serial and apply the workbook's
  default date/time/duration number format when the cell is unstyled
  (I-18/I-19, §7.9). GetDate/GetDateOnly and Kind==Date detect date number
  formats (builtin ids 14-22/45-47 plus custom date-token codes). The
  1900/1904 epochs are honored, including Excel's 1900-02-29 leap bug.
- Applies WorkbookOptions.DefaultFontName/Size to font index 0 on
  CreateOoxml and DateSystem.Excel1904 to workbookPr/@date1904. OpenOoxml
  adopts the file's stylesheet and default font untouched (lessons #8/#9).

Stays within the parallel-engine rules: only Internal/Ooxml*.cs and the
isolated NetXlsx.OoxmlEngine.Tests are touched; the NPOI engine and the
existing suite are unchanged. OoxmlEngine.Tests: 30 -> 52, green on
net8.0 + net10.0.

…-82)

Implements the first half of the sheet/workbook structural surface on the
SDK engine (the slice is split across sessions; panes/visibility/tab color/
protection/grouping follow).

Merges (OoxmlSheet.Merges.cs): <mergeCells> written after <sheetData> in the
worksheet's strict child order; container dropped (not left childless) when the
last region is removed; @count kept in sync. Matches the NPOI engine contract —
1x1 no-op (I-38), overlap throws InvalidOperationException (sec 6.4), non-exact
unmerge is a silent no-op, MergedRanges canonical A1:C3, and MergeCellsStyled
styles every cell before merging so boundary-cell borders render (lesson #4).

Named ranges (OoxmlWorkbook.Names.cs + OoxmlNamedRange.cs): stored as
<definedNames> in workbook.xml between <sheets> and <calcPr>; sheet scope
round-trips via localSheetId. Leading '=' stripped; names unique workbook-wide
case-insensitively regardless of scope; Excel name rules validated in-engine
(the SDK has no built-in validator).

No public symbol added — all are existing interface members newly implemented
internally, so the PublicAPI snapshot is unchanged. Both fixture families added
to the schema-validation gate (clean under Microsoft365). Cross-checked against
the NPOI-engine FreezeMergeHiddenTests (merge half) and NamedRangeApiTests —
every assertion is satisfied, de-risking this surface's cutover.