Add transactional document deletion and replacement

[1816x]

Motivation

• Close the gap in corpus lifecycle management by adding transactional delete and replace operations that preserve existing persistence guarantees.
• Preserve invariants: len(index) == len(chunks), chunk ids dense 0..n-1, every chunk references an existing document, Document.n_chunks matches live chunks, and _next_doc_id never moves backward.
• Use deterministic full HNSW rebuild/compaction (no tombstones) so surviving vectors are reused and chunk ids remain dense and aligned with HNSW ids.

Description

• Add a deterministic internal rebuild primitive DocumentStore._rebuild(...) that orders live vectors by document id ascending then chunk ordinal ascending, reuses survivor vectors via the HNSW vector(id) API, inserts vectors into a fresh HNSW, and regenerates dense StoredChunk metadata and document chunk counts.
• Implement DocumentStore.delete_document(doc_id) and DocumentStore.replace_document(...) with transactional semantics: build staged index/state, call existing _persist when RAG_STATE_PATH is enabled, and only publish the new in-memory state after successful persistence; replacement preserves doc_id and embeds only replacement chunks.
• Surface lifecycle endpoints in the service: PUT /documents/{document_id} and DELETE /documents/{document_id} that reuse DocumentIn validation and RAG_UPLOADS_ENABLED gating, and return proper 404/403/413/422 semantics; update CORS to allow PUT and DELETE without weakening other controls.
• Add same-origin Next.js server helpers (deleteDocument, replaceDocument), a dynamic proxy route app/api/documents/[id]/route.ts, and minimal workspace UI controls for Replace/Delete (confirmation, disabled/loading states, error display, refresh after mutation).
• Add store/regression tests covering deletion, replacement, restart recovery, failed-persistence rollback, dense-id invariants, and embedder call-counting to prove survivors are not re-embedded; update README.md with the rebuild strategy and trade-offs.

Testing

• Ran Rust checks: cargo fmt --all --check (passed), cargo clippy --all-targets -- -D warnings (passed), and cargo test --workspace (passed).
• Built and installed the Python bindings with python -m pip install -e bindings (succeeded) and then ran Python tests: PYTHONPATH=service pytest -q bindings/tests service/tests (73 passed, 1 warning) and targeted lifecycle tests PYTHONPATH=service pytest -q service/tests/test_document_lifecycle.py service/tests/test_e2e.py (38 passed).
• Verified frontend build: cd app && npm ci (succeeded), npm run build (production build succeeded), and npx tsc --noEmit (typecheck passed).
• Additional checks: python -m compileall -q service and git diff --check ran clean.

All automated test runs listed above passed after installing the bindings; initial collection failed before the editable bindings were installed but was resolved by installing bindings and rerunning the tests.