fix(content): restore Content.Ebml build and port VInt tests to xUnit

[dotnetcadet]

Summary

Assimalign.Cohesion.Content.Ebml has not compiled since the 10.0.0 bulk import (e24469bb) — the only commit its folder has ever received. VInt.Read called a ReadFully Stream extension that lives in a different assembly (Content.Media, namespace Assimalign.IO), producing CS1061, and .editorconfig:89 (csharp_prefer_braces = true:error) turned 12 brace omissions in the same file into build errors.

Fixing the build exposed a second, larger problem: the test suite was written for NUnit ([TestCase], Assert.AreEqual, ExpectedResult=) inside a project that references xunit only, so it had never compiled either. The CS1061 was not what kept those tests from running.

Type of change

• fix — bug fix
• test — tests only
• chore — build / tooling / CI

Changes

• VInt.Read: replaced the missing ReadFully with the BCL Stream.ReadExactly (.NET 7+) — precisely "read N bytes or throw EndOfStreamException". Both call sites map to it exactly, including extraBytes == 0 (no-op, no throw). The Content.Media copy uses the legacy this T extension syntax the repo forbids, so it was deliberately not reused; the BCL API also removes the cross-assembly helper dependency that issue [L01.01.05.08.01] Repair EBML VInt, reader, writer, and document lifecycle behavior #459 calls out.
• IDE0011: added braces at all 12 sites. Write's null guard became ArgumentNullException.ThrowIfNull, and the redundant ReferenceEquals(null, obj) in Equals was dropped (obj is VInt already handles null); obj is now correctly object?.
• Tests ported to xUnit + Shouldly, all 40 original cases preserved, with repo-standard [Theory(DisplayName = "Cohesion Test [Content.Ebml] - …")] naming and {Method}_{Scenario}_{ExpectedBehavior} methods. Added the missing Shouldly reference and removed Usings.cs (global usings are forbidden; Content.Text uses explicit ones).
• New coverage for VInt.Read — previously untested, and the one path this fix changes. Includes a ChunkedReadStream fixture that yields one byte per Read call, which is exactly what distinguishes read-exactly from a naive read, plus truncation, empty-stream, over-width, and write/read round-trip cases.
• CI: added Assimalign.Cohesion.Content.Ebml to the library-content.yml matrix.

Work items resolved

Advances #459 — the "consistent exact-read strategy / fails deterministically on truncated input / no reliance on unrelated namespaces" half of its acceptance criteria, plus the VInt encode-decode, truncated-input, and writer-round-trip test coverage.

Not closing it: EbmlReader and EbmlWriter are still empty stubs and EbmlDocument/EbmlBody/EbmlHeader are empty shells, so "reader iteration" and the document lifecycle remain unaddressed.

Discovered (out-of-scope) work

Not fixed here, flagged for follow-up:

• Namespaces do not match assembly names. All 13 files use Assimalign.IO.Ebml, not Assimalign.Cohesion.Content.Ebml (Content.Media has the same problem with Assimalign.IO). Left as-is — renaming is a public-API break that deserves its own decision. Nothing outside each package consumes these namespaces, so the blast radius is small.
• Content.Media still carries the duplicate ReadFully in forbidden legacy this syntax; it should move to Stream.ReadExactly too. Related: [L01.01.05.03] Build shared content stream, buffering, and exact-read primitives #438 (shared exact-read primitives).
• The same bulk-import debt likely affects .Bmff, .Mkv, .Mpeg, .Pdf, .Exe — assume NUnit-era tests that never compiled, so adding them to CI is not a one-line change.
• Pre-existing CS8618 warnings in EbmlDocument.cs are untouched.

Testing & verification

• dotnet build and the affected dotnet test projects pass locally.
• dotnet test libraries/Content/Assimalign.Cohesion.Content.Ebml/tests/ → 60/60 passing, in both Debug and Release (CI builds Release).
• dotnet pack --configuration Release succeeds — CI packs and pushes this project on main.
• The partial-read test was mutation-checked: reverting ReadExactly to a naive Read fails 2 tests, confirming the new coverage is not vacuous.
• ⚠️ Verified on Windows only. The code is pure arithmetic over in-memory buffers — no paths, encodings, or line endings — so cross-platform risk is low, but the ubuntu/macos legs are unconfirmed until this workflow runs.

Checklist

• Deferred/backlog items are linked without a closing keyword.
• Follows the repo coding rules (.claude/rules/): CohesionPackageReference for Shouldly; no Microsoft.Extensions.*; no global usings; Shouldly-only assertions.
• Remains NativeAOT- and trimming-safe — ReadExactly introduces no reflection or dynamic codegen.
• Tests added/updated and passing; the project was already wired into all three .slnx files and is now in the CI workflow.
• No dangling solution/project references left behind.
• Public APIs have XML docs — pre-existing gap in this package (several <summary> blocks are empty); not introduced or widened here.

🤖 Generated with Claude Code