Preserve C# multiline search origin classification - #5316

Summary

Preserve C# multiline comment, verbatim-string and raw-string origins when searching indexed text. The issue's ArgumentList examples no longer pass --origin code. Read bounded indexed prefixes across chunk boundaries, retain source coordinates and schema/help/regex labels, and return unknown when lexical state cannot be established.

CLI rows, filtered counts, token-boundary recipes and MCP share the classifier. No database migration or rebuild is required.

Fixes #5307

Validation

• Build: zero warnings/errors. Search regression suite: 563 passed on net8.0 and 563 passed on net9.0, including rows/counts, recipes, MCP, shared-prefix allocations, cancellation and lexical budgets.
• Real indexing smoke test with 9 chunks: the code filter retains only the executable control; comment and verbatim/raw matches retain lines 303, 406 and 509.
• dotnet run --project tools/CodeIndex.Changelog -- check.
• Codex adversarial review using .codex/workflows/adversarial-review.md (two rounds). Addressed repeated prefix/schema work and interpolation-opener parity, including the second-round finding for many distinct matched literals. Allocation regressions cover both code matches and 500 string matches; focused regressions were rerun after the final shared-text validation change.

Documentation

Updated English and Japanese README, developer-guide and testing-guide sections. Added changelog.d/unreleased/5307.fixed.md; CHANGELOG.md is unchanged. Agent entry points and AGENT_GUIDE.md are unchanged.

Follow-ups

None outside this issue. Missing/budget-limited prefixes and text at/after interpolation expressions retain the documented unknown origin; inspect those results without a code-origin filter.