Reduce allocations in template matching tries

[JeremyKuhne]

Summary

• store the common single-child trie case inline
• materialize the shipped NextNodes dictionary only when a node branches or callers request it
• keep evaluator lookups on the compact representation without changing matching behavior
• test that NextNodes remains stable and reflects paths added after materialization

Performance

Measured with matched Windows x64 Release ReadyToRun SDK trees differing only in Microsoft.TemplateEngine.Core.dll.

• reported allocation: 49,401,248 B to 40,047,344 B (18.93% reduction)
• allocation events: 293 to 205
• peak heap: 33.87 MB to 25.05 MB
• initialized console creation, 30 alternating pairs: effectively neutral (-0.17% paired median, 15/30 pairs faster)
• initialized template list: 4.08% faster
• template help: 1.75% faster
• initialized class library creation: 3.35% faster

All 80 semantic comparisons matched across 160 fresh process launches.

Validation

• Microsoft.TemplateEngine.Core.UnitTests, Debug and Release, net11.0
• Microsoft.TemplateEngine.Core.UnitTests, Debug and Release, net481
• 363 passed, 1 existing skip, 0 failed in each captured full run
• full Release ReadyToRun SDK build: 0 warnings, 0 errors

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).
2 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟡 Changes recommended

The newly added unit test likely fails to compile under repo-wide nullable+warnings-as-errors due to dereferencing nullable out variables without proving non-null to flow analysis.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR optimizes the TemplateEngine matching trie to reduce allocations by avoiding eager Dictionary creation for the common single-child case, while preserving matching behavior and providing a stability contract for NextNodes once materialized.

Changes:

• Add a compact single-child representation to Trie<T> and lazily materialize the NextNodes dictionary only when needed.
• Update TrieEvaluator<T> to traverse using the new compact TryGetNextNode lookup to avoid forcing dictionary allocation.
• Add a unit test validating NextNodes materialization stability and that subsequent additions are reflected.

File summaries

File	Description
test/TemplateEngine/Microsoft.TemplateEngine.Core.UnitTests/TokenTrieTests.cs	Adds coverage for NextNodes stability/materialization behavior.
src/TemplateEngine/Microsoft.TemplateEngine.Core/Matching/TrieEvaluator.cs	Switches evaluator traversal from NextNodes.TryGetValue to TryGetNextNode.
src/TemplateEngine/Microsoft.TemplateEngine.Core/Matching/Trie.cs	Implements compact single-child storage and lazy dictionary materialization with a new lookup API.

Review details

• Files reviewed: 3/3 changed files
• Comments generated: 1
• Review effort level: Lite

[JeremyKuhne]

The failing CI result is unrelated to this PR. The only failing test is ScopedCss_PublishIsBackwardsCompatible_WithPreviousVersions, where Razor exits with This writer does not support components. This exact regression is tracked by dotnet/roslyn#85132, and the corresponding SDK test quarantine is dotnet/sdk#56134. All platform builds, AOT builds, and the Windows Full Framework leg passed.

[JeremyKuhne]

The remaining CI failure is tracked by #56238. The Linux x64 -mt row failed with MSB3883: Cannot release a lock that is not owned by the current thread; the serial row and recent PR/main executions passed.

[JeremyKuhne]

Build 20260910.31 failed because of two environment-variable races exposed by the class-level test parallelization added in #56036. Neither intersects this PR's Template Engine trie changes.

• CA CLI invocation tests flake on Mac arm64 Helix due to -tlp:DISABLENODEDISPLAY leaking through AdjustMSBuildForLLMs #56258 tracks the unexpected -tlp:DISABLENODEDISPLAY argument. This run's shard included the LLM-environment mutator and GivenDotnetCleanInvocation; only one clean-command row failed, identifying transient in-process state rather than a stable Helix setting.
• [Flaky test] MTPHelpSnapshotTests races with DOTNET_NOLOGO mutation under parallel execution #56264 tracks MTPHelpSnapshotTests, where transient DOTNET_NOLOGO=true changed the help snapshot default from False to True.

#56036 added ResourceLock(EnvironmentVariables) to the mutating tests, but the lock only coordinates with readers declaring the same resource. Both failed readers remain unannotated. All individual platform build legs passed; the four red checks are summaries of these two Helix failures.