Fix BinLogReader for Linux binlogs and HtmlGenerator duplicate serverPath

[ericstj]

Summary

Fix BinLogToSln/HtmlGenerator tool issues discovered while adding VMR source-index support.

1. BinLogReader: Use tree-based reader instead of event-based Replay()

BinLogReader.Replay() silently aborts when reading binlogs that lack a CurrentUICulture message — a known issue (MSBuildStructuredLog#936). This is common in binlogs produced on Linux. When Replay() hits a PropertyReassignment record, it tries to format a resource string that was never initialized, throws an ArgumentNullException, catches it internally via OnException, and stops replaying — with zero diagnostics to the caller.

Impact for VMR source-build: BinLogToSln extracted zero compiler invocations from Linux source-build binlogs (arcade, runtime, roslyn, etc.), producing empty SLN files. The same binlogs processed with the tree-based Serialization.Read() API work perfectly — arcade yields 25 invocations, runtime yields 895.

Impact for existing per-repo pipelines: Most repos (runtime, roslyn, aspnetcore) run their source-index job on Windows agents or cross-compile with -os linux on Windows, so their binlogs likely include the culture message and are unaffected. However, any repo that runs source-index natively on Linux could be silently producing empty results. Worth auditing.

The fix switches ExtractInvocations to always use the tree-based Serialization.Read() path (which was already implemented for .buildlog files). This approach:

• Works reliably for all binlog formats (no Strings.Initialize() workaround needed)
• Is simpler — no event wiring or state tracking dictionaries
• Throws on parse errors instead of silently aborting
• Fully populates ProjectProperties via GetEvaluation(build).GetProperties() (verified: ~1000 properties per invocation including TargetFramework)

Also adds Linux compiler path trimming — on Linux, the Csc task command line starts with /path/to/Roslyn/bincore/csc (no .exe extension), which the old trimming logic didn't handle.

A workaround exists (Strings.Initialize() before Replay(), as used in component-detection), but the tree-based approach is strictly better for our use case.

2. HtmlGenerator: Allow duplicate serverPath mappings

When multiple SLN files from the same VMR build map to the same source directory (e.g., dotnet-dotnet and dotnet-dotnet-windows both mapping to src/), HtmlGenerator crashed with Dictionary.Add on duplicate keys. Changed to indexer assignment to allow last-write-wins.

Testing

• Verified against actual Linux source-build binlogs downloaded from VMR build 2973064
• All 10 existing BinLogToSln tests pass
• Arcade: 0 → 25 invocations, Runtime: 0 → 895 invocations
• Windows binlogs (winforms): 29 invocations (unchanged, still works)
• Property extraction verified: all invocations have full properties including TargetFramework

[ericstj]

I've validated these locally and in the pipeline. I need to get these merged and consumed in VMR to get a good set of data out of VMR. Right now it's busted and likely has been busted on linux for a while, we just didn't happen to have any pipelines building on linux for their source index data.

[ericstj]

https://dev.azure.com/dnceng/internal/_build/results?buildId=2975092&view=results