Fix Anthropic replay normalization for adjacent text blocks

[stephentoub]

Summary

• canonicalize adjacent plain-text Anthropic user content blocks with the replay harness's existing \n\n\n user-message separator
• share the canonical separator with request and snapshot coalescing so those normalization paths cannot drift
• add adapter and protocol-aware replay regressions while retaining single-block, multimodal, and tool-result coverage

Root cause

github/copilot-agent-runtime#15832 intentionally preserves adjacent user messages as separate Anthropic content blocks for prompt-cache stability. The replay adapter flattened those block boundaries with \n, while captured canonical snapshots and adjacent-user-message coalescing use \n\n\n. Requests therefore failed to match snapshots during compaction and session-limit flows.

This change fixes normalization only at the replay adapter boundary; it does not alter the runtime's Anthropic serialization or broadly relax matching.

Tests

• cd test/harness && npm test (61 passed)
• COPILOT_SDK_E2E_BACKEND=anthropic-messages COPILOT_SDK_DEFAULT_CONNECTION=inprocess dotnet test test/GitHub.Copilot.SDK.Test.csproj -f net8.0 --no-restore -v normal --filter FullyQualifiedName=GitHub.Copilot.Test.E2E.CompactionE2ETests.Should_Trigger_Compaction_With_Low_Threshold_And_Emit_Events

[Copilot]

Pull request overview

Aligns Anthropic replay normalization with canonical snapshot coalescing.

Changes:

• Shares the canonical user-message separator.
• Normalizes adjacent Anthropic text blocks.
• Adds adapter and protocol replay regressions.

Show a summary per file

File	Description
test/harness/replayingCapiProxy.ts	Reuses the shared separator.
test/harness/modelProtocolAdapterShared.ts	Defines the canonical separator.
test/harness/modelProtocolAdapters.test.ts	Adds normalization regressions.
test/harness/anthropicMessagesAdapter.ts	Joins adjacent text blocks canonically.

Review details

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

• Files reviewed: 4/4 changed files
• Comments generated: 1
• Review effort level: Balanced