fix: data integrity, route matching, namespace filter, vector growth

[ajianaz]

What

Empat perbaikan P1 dari audit cora-code v0.13.0:

1. crud.rs atomicity — insert() sekarang wrap INSERT + tag inserts dalam unchecked_transaction(), konsisten dengan pattern di tags.rs dan documents.rs. Jika tag insert gagal, seluruh operasi rollback.

2. handlers.rs route matching — Query string sekarang di-strip sebelum route dispatch (route_path = path.split("?").next()). Semua handler yang extract query params sekarang pakai path.split("?") bukan p.strip_prefix(...). Menghilangkan 14 clippy redundant_guard warnings.

3. lib.rs namespace filter — graph_data(namespace) sebelumnya no-op (is_none_or(|_| true) selalu return true). Sekarang query memory IDs by namespace, filter nodes by memory_id, dan filter edges untuk consistency.

4. vector.rs capacity growth — Fixed increment +1024 → geometric growth max(current*2, current+4096, 1024). Amortizes to O(1) per insertion untuk dataset besar.

Why

• crud.rs: Crash di tengah tag inserts meninggalkan memory tanpa tags → data inconsistency
• handlers.rs: Route matching dengan query string bikin fragile dan redundant guard warnings
• lib.rs: Namespace filter yang no-op = fitur tidak berfungsi, user tidak bisa filter graph by namespace
• vector.rs: Fixed increment menyebabkan frequent reallocation pada dataset 100K+ vectors

Changes

• crates/uteke-core/src/memory/crud.rs — unchecked_transaction() wrapper
• crates/uteke-server/src/handlers.rs — route_path variable + literal path matching
• crates/uteke-core/src/lib.rs — Namespace-aware graph node/edge filtering
• crates/uteke-core/src/memory/vector.rs — Geometric capacity growth formula

Testing

• 476 tests pass, 0 fail, 25 ignored
• Clippy: 0 warnings
• Cora review: No issues found

[github-actions]

🔍 Cora AI Code Review

❌ Blocked — critical issues found.

🔴 Error (2)

• crates/uteke-core/src/lib.rs:1160 — The new namespace filtering logic executes a SELECT id FROM memories WHERE namespace = ?1 AND deprecated = 0 query and collects all results into a HashSet<String> on every call to graph_data when a namespace is specified. For databases with a large number of memories in a namespace, this loads potentially thousands of strings into memory just to perform a set membership test for graph node filtering. This is an O(N) memory allocation where N is the total memory count in the namespace, even if the graph only contains a handful of nodes.
• crates/uteke-core/src/memory/crud.rs:80 — The comment states unchecked_transaction is "safe because Store access is serialized via Mutex" — however, unchecked_transaction bypasses rusqlite's compile-time safety checks for transaction nesting and does not actually lock the connection. If any other code path calls methods on this Store using a shared &Connection (even read-only methods like prepare/query), those operations will interleave with this transaction and potentially see partial state or cause SQLITE_BUSY errors. The safety depends entirely on the caller holding the Mutex for the entire duration, but Store methods take &self which doesn't enforce this invariant.

---

Review powered by cora-code · BYOK · MIT