fix: hash reference lookup failures with Codex models

Four fixes for "Hash reference not found" errors:

1. Normalize hash refs — trim whitespace, strip trailing "|",
   handle spaces around ":" (models may format refs loosely)

2. Line truncation consistency — computeFileHashes now truncates
   lines >2000 chars before hashing, matching OpenCode's read tool
   output. Previously produced different hashes for long lines.

3. Recompute hashes after edits — instead of clearing, recompute
   from the new file content so subsequent stale-ref errors show
   "line N now has hash X" instead of generic "not found"

4. Diagnostic error messages — when hash isn't found, shows what
   hash IS at that line number (or that the line doesn't exist),
   helping the model recover by re-reading

Also adds explicit system prompt guidance on hash ref format
("copy exactly, don't include | separator").

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>