perf(server): cache HTTP query embeddings

[justonemorenight]

What

Install the existing CachingEmbedder on the long-lived HTTP query path that #2189 explicitly left as a follow-up.

The wrapper is applied both to the primary HTTP vector store and to lazy workspace repo contexts. This keeps one cache per embedder instance/configuration and avoids repeated provider calls for repeated searches.

Safety

• Leaves KeylessEmbedder bare so is_semantic_embedder() continues to disable the signal-free vector leg.
• Reuses the bounded, copy-on-read LRU from perf(mcp): cache query embeddings so a repeated search skips the provider #2189; no new cache policy or persistence.
• Keeps _build_embedder() returning the concrete backend for callers that need backend-specific attributes.
• Bulk embedding remains pass-through, so indexing batches do not fill the query cache.

Tests

• Repeated HTTP query text reaches the underlying embedder once.
• Keyless HTTP embedder is not wrapped.
• uv run ruff check packages/server/src/repowise/server/app.py tests/unit/server/test_app_embedder_backends.py
• uv run pytest tests/unit/server/test_app_embedder_backends.py tests/unit/test_persistence/test_caching_embedder.py -q (14 passed)
• uv run pytest tests/unit/server/test_app_workspace_registry.py tests/unit/server/test_jobs.py -q (10 passed)
• uv run pytest tests/unit/server --ignore=tests/unit/server/mcp -q (1323 passed)

[RaghavChamadiya]

Leaving KeylessEmbedder bare is the detail that makes this correct, and the reason is worth the two lines you gave it: is_semantic_embedder fails open by design (base.py:185), so a wrapped keyless would come back True and switch on a vector leg with no signal. Checking the raw embedder before wrapping keeps both answers right. One thing to keep an eye on rather than change here: embedder_factory is called once per workspace member (registry.py:242), so a large workspace now holds one 256-entry LRU per repo instead of one. Bounded and only filled by real queries, so not a blocker, but it is worth remembering if a workspace ever reports memory growth on the serve path.