Add function-level merge fallback for whole-file merge failures

[TheValiantOne]

Summary

DiffPlexMergeEngine.MergeHeadless merges whole .ws files line-by-line, which means real, incidental whitespace/comment differences (or a confirmed upstream DiffPlex bug that gets worse as edit density rises — see WitcherScriptMerger.Core/CLAUDE.md's "Compatibility constraint" section) can block an entire file from auto-solving even when the actual overlapping logic changes are confined to one or two functions.

This adds a fallback, activated only at the two points where the whole-file merge has already failed for a given pairwise chain step — never a parallel code path, so every conflict that already auto-solves today is unaffected:

• ScriptUnitExtractor — splits a .ws file into function/event/@addField-field units via a string/comment-aware brace-matching tokenizer. Not a full parser: confirmed via direct research into WitcherScript's grammar that class/state/struct/enum declarations are top-level only and the language has no nested function-like constructs at all, so plain brace/paren counting is sufficient.
• UnitAligner — aligns each side's units against vanilla's by name (LCS-based), handling insertions and deletions symmetrically.
• FunctionLevelMergeEngine — resolves each vanilla function independently: cheap shortcuts for untouched/single-sided edits, a real per-function 3-way merge for non-overlapping changes, and a most-distinct-from-vanilla tiebreak (via DiffPlex.Differ's 2-way line diff, not the buggy ThreeWayDiffer) for genuine collisions. An edit always wins over a competing deletion. Every non-mechanical resolution is recorded in an audit trail — surfaced via DiffPlexMergeEngine.LastFunctionLevelDecisions → FileMerger.HeadlessMergeSummary.FunctionLevelDecisions → both hosts' CLI output and the MCP merge_conflicts tool's new functionLevelDecisions field — never applied silently.

Full design rationale is in WitcherScriptMerger.Core/CLAUDE.md's new "Function-level merge engine" section.

AI-assisted development disclosure (per CONTRIBUTING.md): this PR was developed with Claude Code, including the design research, empirical validation, and implementation.

Why this design

Before writing any code, I measured a real, live Witcher 3 install's actor.ws conflict (vanilla + 6 real overhaul mods): only 6 of 395 functions were genuine two-mod collisions, confirming per-function decomposition was worth building rather than just relocating the same conflict into a smaller, statistically more DiffAlgorithmException-prone box.

Mid-implementation, a chain-step replay against the real install's 5 currently-unresolved conflicts found that an insertion-only alignment (the originally-scoped design) would rescue only 2 of the 5 — one real mod in that install deletes several vanilla functions outright, and that deletion was found to persist through the merge chain even at steps where the deleting mod isn't a direct input (an earlier clean whole-file merge step faithfully propagates a one-sided deletion into the accumulated text). UnitAligner's symmetric insert/delete handling exists specifically to cover this — confirmed as the right call before building it further.

Test plan

• dotnet build WitcherScriptMerger.sln --configuration Release — clean, 0 errors (pre-existing unrelated CA1823 warnings only)
• dotnet format whitespace WitcherScriptMerger.sln --verify-no-changes — clean
• dotnet test WitcherScriptMerger.sln — 125/125 passing, including 3 new test files (ScriptUnitExtractorTests, UnitAlignerTests, FunctionLevelMergeEngineTests) covering round-trip extraction fidelity, alignment (matches/insertions/deletions/both), every one-sided shortcut, the tiebreak (including its deterministic tie-break and a scaled-down DiffAlgorithmException case), edit-survives-competing-deletion, insertion reconciliation, and gap-comment detection
• Real-data validation, not just synthetic fixtures: built an isolated scratch copy of a real Witcher 3 install's vanilla files + real mod files for the 5 conflicts that don't auto-solve today, ran the actual MCP merge_conflicts tool against it end-to-end. All 5 files merged successfully, including actor.ws itself (the flagship case that motivated this feature) — well-formed UTF-16LE+BOM output, no conflict markers, itemized functionLevelDecisions explaining every resolution. Never run against the live install directly (dry-run predicted the "output already exists" guard correctly with no writes).

🤖 Generated with Claude Code

https://claude.ai/code/session_01GXAuGMLB44T5Zv5o5ZzKah