feat: LRU embedding cache for workspace search (#1423)

* feat: LRU embedding cache for workspace search (#165)

Add CachedEmbeddingProvider that wraps any EmbeddingProvider with an
in-memory LRU cache keyed by SHA-256(model_name + text). This avoids
redundant HTTP calls when the same text is embedded multiple times
(common during reindexing and repeated searches).

- Cache uses HashMap + last_accessed tracking with manual LRU eviction
  (same pattern as llm::response_cache::CachedProvider)
- Lock is never held during HTTP calls to prevent blocking
- embed_batch() partitions into hits/misses and only fetches misses
- Default 10,000 entries (~58 MB for 1536-dim vectors)
- Configurable via EMBEDDING_CACHE_SIZE env var
- Workspace.with_embeddings() auto-wraps; with_embeddings_uncached()
  available for tests

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address review comments on embedding cache

- Validate embed_batch return count matches expected miss count
- Replace unwrap_or_default() with proper error propagation
- Fix batch eviction: run final eviction pass after insert to enforce cap
- Fix test: use different-length inputs to verify ordering correctness
- Reject EMBEDDING_CACHE_SIZE=0 in config validation (minimum is 1)

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: replace .expect() with proper error handling in embed_batch

The all-cache-hits early-return path used .expect("all cache hits") which
violates the project convention of no .unwrap()/.expect() in production
code. Replaced with the same ok_or_else pattern used in the normal path.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: clarify memory sizing docs and use saturating_add for eviction

- Update memory comments in embedding_cache.rs, config/embeddings.rs,
  and workspace/mod.rs to note the ~58 MB figure is payload-only
  (actual memory is higher due to HashMap/key/allocation overhead)
- Use saturating_add(1) instead of + 1 for eviction threshold to
  prevent overflow if max_entries is usize::MAX

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address Copilot review on embedding cache

- Avoid double-clone per miss in embed_batch: move embedding into
  results, clone only for the cache entry
- Evict per-insert instead of after all inserts to keep peak memory
  bounded during large batches
- Clamp max_entries to at least 1 in constructor to prevent unexpected
  eviction behavior when set to 0 via the public API

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: reduce embedding_cache module visibility to private

Types are already re-exported via `pub use`, so the module itself
doesn't need to be public. Reduces unnecessary API surface.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address serrrfirat review feedback on embedding cache

- Add TODO comment for O(n) LRU eviction scalability
- Add thundering herd note at lock release in embed()
- Warn when cache max_entries exceeds 100k
- Use with_embeddings_uncached() in integration test
- Add tests: error_does_not_pollute_cache, embed_batch_empty_input
- Update README with cache-aware with_embeddings() docs

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: prevent u32 wrapping in FailThenSucceedMock failure counter

fetch_sub(1) wraps to u32::MAX when called past zero, silently
breaking the mock for 3+ calls. Switch to load-then-store to avoid
the wrapping bug in both embed() and embed_batch().

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address Copilot and serrrfirat review findings on embedding cache

- Switch tokio::sync::Mutex to std::sync::Mutex (lock never held across
  .await — cheaper synchronous lock)
- Extract DEFAULT_EMBEDDING_CACHE_SIZE constant to avoid 10_000 duplication
  between EmbeddingCacheConfig and EmbeddingsConfig

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* test: add all-misses batch test for embedding cache

Adds embed_batch_all_misses test covering the case where every text in a
batch is a cache miss — fulfilling the commitment from serrrfirat's review.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* chore: trigger CI re-check after rebase

* fix: use raw [u8;32] cache keys and pre-allocate HashMap capacity

Address Copilot review findings:
- cache_key() now returns [u8; 32] instead of hex String, avoiding a
  64-byte allocation per lookup
- HashMap::with_capacity(max_entries) avoids incremental reallocation
- Fix pre-existing staging compilation error in cli/routines.rs
  (missing max_tool_rounds/use_tools fields)

[skip-regression-check]

* fix: make cache accessors sync and update doc for [u8;32] keys

Address Copilot review:
- len(), is_empty(), clear() are now sync since they only take a
  std::sync::Mutex lock with no .await points
- Update cache_size doc comment to reflect [u8;32] keys instead of
  String keys

[skip-regression-check]

* fix: remove clone_on_copy for [u8; 32] cache keys

[skip-regression-check]

* ci: add safety comments to test code for no-panics check

The CI no-panics grep check cannot distinguish test code inside
src/ files from production code. Add // safety: test annotations
to .unwrap(), .expect(), and assert!() calls in #[cfg(test)] modules.

* fix: correct cache doc and demote hit/miss logs to trace

- Fix misleading "String keys" in memory comment (cache uses [u8; 32])
- Demote per-request hit/miss logs from debug to trace to reduce noise
  on hot paths (batch summary stays at trace too)

* docs: add missing Arc import in workspace README example

* perf: batch eviction in embed_batch to avoid O(n×m) cost

Replace per-insert evict_lru call with a single evict_k_oldest pass
that computes eviction count upfront and removes the k oldest entries
in one O(n) scan. Avoids O(n×m) HashMap iterations while holding the
mutex during batch inserts.

* fix: cap batch cache inserts at max_entries and use O(n) selection

- evict_k_oldest now uses select_nth_unstable_by_key for O(n) average
  partial selection instead of O(n log n) full sort
- embed_batch caps cached entries at max_entries when misses exceed
  capacity, preventing the cache from growing unbounded
- Added test: batch_exceeding_capacity_respects_max_entries

* fix: flatten test assert for fmt compatibility

Shorten assert message to fit single line so cargo fmt doesn't
split the safety annotation onto a separate line.

* fix: address review feedback and improve embedding cache (takeover #235)

- Fix merge conflict: add missing allow_always field in PendingApproval
- Thread EmbeddingCacheConfig through CLI memory commands so they respect
  EMBEDDING_CACHE_SIZE instead of silently using default (fixes #235 review)
- Cap HashMap pre-allocation at min(max_entries, 1024) to avoid upfront
  memory waste at large cache sizes
- Fix FailThenSucceedMock race: replace load+store with atomic fetch_update
- Remove noisy '// safety: test' comments (40+ lines of diff noise)
- Fix collapsed lines from comment removal
- Simplify redundant Ok(...collect()?) to just collect()

Co-Authored-By: ztsalexey <[email protected]>
Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* fix(embedding-cache): skip eviction on concurrent duplicate insert

When the lock is released for the HTTP call, another caller may insert
the same key. Re-check under lock and just update the existing entry
without evicting, avoiding unnecessary cache churn under concurrency.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

---------

Co-authored-by: ztsalexey <[email protected]>
Co-authored-by: Claude Opus 4.6 <[email protected]>
Co-authored-by: ztsalexey <[email protected]>

tests/workspace_integration.rs

@@ -308,7 +308,7 @@ async fn test_workspace_hybrid_search_with_mock_embeddings() {
 
     // Create workspace with mock embeddings (1536 dimensions to match OpenAI)
     let embeddings = Arc::new(MockEmbeddings::new(1536));
-    let workspace = Workspace::new(user_id, pool.clone()).with_embeddings(embeddings);
+    let workspace = Workspace::new(user_id, pool.clone()).with_embeddings_uncached(embeddings);
 
     // Write documents
     workspace