Fix correctness bugs found by review in the function-level merge engine

[TheValiantOne]

Summary

A multi-angle code review of the just-merged function-level merge engine (#18) surfaced several real bugs, all fixed here:

• stdout corruption (critical for MCP): TryFunctionLevelRescue's success notification used DialogIcon.Information, which HeadlessMergeNotifier routes to stdout — corrupting the MCP JSON-RPC stream on every rescue that had decisions to report. Now DialogIcon.Warning (stderr), matching every other message in the class.
• Crash on zero-function files: a .ws file with no extracted functions/fields (e.g. only top-level consts/enums) crashed with IndexOutOfRangeException in GetSideGapIndex, silently swallowed by a bare catch — the rescue just never fired for exactly this file shape. Fixed the index bug, and separately made TryMerge decline outright when vanilla has zero units: the whole document is one "gap" in that case, and gap content always reverts to vanilla on reassembly — so "succeeding" there would have silently discarded both mods' real edits instead of the safe decline this now is.
• Crash on duplicate insertion names: ReconcileInsertions threw via ToDictionary when one side inserted two same-named units at the same slot — now declines cleanly.
• NBSP regression: NormalizeGap used plain Trim() instead of the NBSP-safe Trim(WhitespaceChars), reintroducing a bug DiffPlexMergeEngine.NormalizeWhitespace was already fixed for. Now reuses that method directly (promoted to internal) instead of a drifted second copy.
• Silent content loss: a gap slot made "ineligible" for precise comparison by a nearby deletion got zero signal even when its content actually differed — contradicting the engine's own "never empty content silently" contract. Now emits a conservative, location-described caveat note.
• Missing description threading: MergeBundleConflictHeadless and MergeTextInteractive (both flat and bundle) never threaded the accumulated-mod-names description MergeFlatConflictHeadless already built, so a bundle-packed or interactively-merged multi-mod chain got a misleading "kept MergedBundleContent's edit" instead of naming the real mods — including in a real GUI modal on the interactive path. All four call sites now share one DescribeAccumulated helper.
• Decisions misattributed to skipped files: function-level decisions recorded during an earlier, successful chain step still got reported even when a later step failed (or a bundle repack failed) and the file ended up in summary.Skipped, not summary.Merged — plus a casing mismatch between merge.RelativePath and conflict.RelativePath could break exact-match correlation downstream. Decisions are now tracked per relative path and only drained for paths that survived to summary.Merged.
• Silent bug-swallowing: the bare catch around the new engine swallowed every exception with zero trace, making a genuine bug indistinguishable from an expected decline. Narrowed to the one anticipated exception type; anything else is now reported via the notifier instead of vanishing silently.
• Performance: ScriptUnitExtractor's @addField lookup rescanned from the current position to EOF on every extracted unit — for a large vanilla file with zero @addField annotations (a mod-only construct), that's millions of wasted line scans per rescue attempt. Now a single forward pass computed once, walked with a monotonic pointer.
• Docs: fixed a stale CLAUDE.md claim that ScriptUnitExtractor reuses FileEncoding.cs for file I/O — it takes pre-read strings and does none.

Not changed (flagged, not acted on): two review angles raised a philosophical concern about the "most distinct from vanilla wins" tiebreak silently resolving genuine edit-vs-edit collisions with no separate review tier. That's the explicitly-requested, approved design from #18, not a bug — left as-is.

AI-assisted development disclosure (per CONTRIBUTING.md): this PR, including the review that found these issues and the fixes themselves, was developed with Claude Code.

Test plan

• dotnet build WitcherScriptMerger.sln --configuration Release — clean
• dotnet format whitespace WitcherScriptMerger.sln --verify-no-changes — clean
• dotnet test WitcherScriptMerger.sln — 128/128 passing, including 4 new regression tests (zero-vanilla-units decline, duplicate-insertion-name decline, plus updated assertions on 3 existing tests to cover the new ineligible-deletion caveat notes)
• Real-data re-validation: re-ran the same isolated-scratch-copy end-to-end check used to validate Add function-level merge fallback for whole-file merge failures #18 (real vanilla + real mod files for the 5 conflicts that don't auto-solve today) — all 5 files, including actor.ws, still merge successfully after these fixes, with well-formed UTF-16LE+BOM output and no conflict markers.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GXAuGMLB44T5Zv5o5ZzKah