[SemutJahat]

Summary

Fixes several problems in how rag_index_project keeps the index in sync with the code, and makes agents depend less on calling it after every task.

Bugs fixed

• Stale chunks after editing a file. A sync only deleted chunks whose whole file was gone. When a file shrank (e.g. functions removed), its old chunks stayed and were served to agents as current code. A sync now deletes every chunk of the directory that the scan did not produce again.
• Re-embedding everything below an edit. Chunk IDs were positional (#chunkN), so inserting a function shifted the ID of every chunk after it. IDs now derive from the chunk's content (dir/file#<hash>, chunk_version v4), and only small declarations are packed together, so an edit only touches nearby chunks. A chunk that only moved gets its line numbers refreshed without re-embedding on SQLite (new optional rag.PayloadUpdater), and is re-indexed on other stores.
• pgvector: ListPoints ignores the source_dir filter, so a sync could delete another directory's chunks. The indexer now only deletes chunks under the directory's own ID prefix.
• Chroma: Index used /add, which silently skips IDs that already exist, so edited chunks were never stored. It now uses /upsert.
• Missing directory wiped the index. rag_index_project on a path that doesn't exist (a typo, a moved repo) did an empty scan and deleted everything. It now returns an error.
• Unreadable file. A file that exists but can't be read keeps its chunks instead of losing them to the stale cleanup.
• Same-named directories. A second directory with the same folder name in one project (/x/app, /y/app) is refused, since their chunks share an ID prefix and would replace each other.

Improvements

• Big classes split at their methods. A declaration too long for one chunk is split at its members (Python, Rust, Ruby, Java/C#, Kotlin, Swift, PHP, JS/TS class methods), with decorators/annotations kept with each method and closures left inside it. Symbols look like Class.method. Before, these were cut at blank lines.
• Auto-sync before search. A directory indexed once with rag_index_project is recorded in ~/.enowx-rag/project_dirs.json and re-synced incrementally before a search when its index is older than RAG_AUTO_SYNC_INTERVAL (default 60s, off disables). Missing directories and projects deleted elsewhere are skipped. The registry file is safe to share between processes. The agent templates (skill, skill/templates/*, write-agents-md, docs) no longer tell agents to run rag_index_project after every task.
• Personal paths (/Users/...) and a personal deployment name removed from README, AGENTS.md, CLAUDE.md and the skill.

Behavior changes

• One-time re-embed: chunk IDs changed, so the first sync after upgrading re-embeds each indexed project once. Old-format chunks are cleaned up automatically.
• More, smaller chunks: functions of fair size now get a chunk each instead of being packed together.
• SyncResult gains updated, and the rag_index_project tool returns chunks_updated.

Testing

• 25 new tests: sync regressions (shrinking file, insert at top, legacy IDs and notes, missing directory, duplicate chunks, provider ignoring the dir filter, unreadable file, no-op second sync, fallback without PayloadUpdater), member splitting per language, SQLite UpdatePayload, and auto-sync (edits seen without re-index, interval/off, missing dir, deleted project not recreated, same-name refusal, shared registry, interval parsing).
• The new regression tests fail when their fix is reverted.
• pkg/rag, pkg/indexer and pkg/core pass go test -race and go vet locally on Go 1.24, using a test-only SQLite driver build because the Go module proxy wasn't reachable from the dev environment.
• The changes in cmd/mcp-server, pkg/config and pkg/httpapi (wiring, one path helper, template text) could not be compiled locally and were reviewed by hand. CI on this PR is their first real build with Go 1.26.
• Not run against live Qdrant, Chroma or pgvector servers.