revert: remove retrieval-timeout optimization series

CONTEXT.md

@@ -173,76 +173,6 @@ The query workflow that returns cited evidence from published documents.
 The typed retrieval request that owns cache-shaping fields and route policy:
 scope, filters, data type, channels, ranking options, and agentic toggle.
 
-### Retrieval Run
-
-One execution of a Retrieval Query, recorded across classic, map-nav,
-small-corpus, cache-hit, failed, and cancelled outcomes with route, timing,
-and terminal-status metadata.
-
-### Retrieval Progress Event
-
-A safe, user-facing update about the current phase of a Retrieval run. It uses
-the fixed phases `started`, `planning`, `searching`, `reviewing_sources`, and
-`finalizing`; it never contains chain-of-thought or raw planner output.
-
-### Retrieval Stream
-
-The server-to-client event stream for a Retrieval Query. It carries
-Retrieval Progress Events during execution and one authoritative final result
-or terminal failure, while leaving answer generation to downstream clients.
-
-### Retrieval Duration
-
-The end-to-end server time for a Retrieval Query, measured from retrieval
-execution start through final public-result assembly. It includes cache lookup
-and excludes authentication, network transfer, SSE delivery time, and
-downstream answer generation.
-
-### Retrieval Non-LLM Work
-
-The database and retrieval-engine work for a Retrieval Query: snapshot or
-serving-index loading, lexical scoring, ranking, result hydration, citation
-assembly, and asset-reference resolution. It excludes planner, harvest,
-control, and answer-generation model time, which are measured separately.
-
-### Retrieval Serving Index
-
-The publication-derived read model used to load retrieval structure and
-scoring inputs without rebuilding them from the full document corpus for each
-query. It is revision-pinned and complete before its document revision becomes
-active.
-
-### Retrieval Serving Fallback
-
-The exact legacy retrieval path used when a serving index is missing,
-incomplete, or inconsistent. It preserves retrieval quality while sacrificing
-the serving-index latency target until the derived data is repaired.
-
-### Retrieval Serving Generation
-
-The namespace-scoped version that identifies one coherent set of active
-document revisions and their serving-index statistics. Retrieval captures one
-generation and retries or falls back if publication changes it during capture.
-
-### Retrieval Semantic Parity
-
-The compatibility requirement that a serving-index retrieval returns the same
-selected chunk IDs, ordering, rounded scores, citations, and asset references
-as the legacy retrieval path for the same request.
-
-### Retrieval Revision Pin
-
-The set of document revision IDs captured at retrieval start and used for the
-entire retrieval run, including lazy content and asset resolution. A later
-publication affects subsequent runs, not the run already in progress.
-
-### Online Retrieval Serving Rollout
-
-The additive rollout of retrieval-serving schema and derived data while
-retrieval and document publication remain available. Incomplete or
-inconsistent revisions use the exact legacy retrieval path until backfill and
-validation finish.
-
 ### Workflow Run Request
 
 The agentic Retrieval request passed through planning and step execution. It

apps/api/alembic/env.py

@@ -101,10 +101,6 @@ def run_migrations_offline() -> None:
         include_object=include_object,
         literal_binds=True,
         dialect_opts={"paramstyle": "named"},
-        # Keep each migration in its own transaction so migrations that
-        # require an autocommit block (for example CREATE INDEX
-        # CONCURRENTLY) can safely commit only their own predecessor.
-        transaction_per_migration=True,
         # Pass through configured SSL connect args.
         connect_args=ssl_connect_args,
     )
@@ -123,25 +119,13 @@ def run_migrations_online() -> None:
     configured_connection = config.attributes.get("connection")
 
     def run_with_connection(connection: Connection) -> None:
-        caller_owned_transaction = connection.in_transaction()
         if settings.API_STANDALONE_MODE_ENABLED:
             ensure_better_auth_user_table(connection)
-            # The standalone bootstrap query starts SQLAlchemy's implicit
-            # transaction before Alembic begins tracking migration
-            # transactions.  End only that transaction; never commit a
-            # transaction supplied by the caller.
-            if not caller_owned_transaction:
-                connection.commit()
 
         context.configure(
             connection=connection,
             target_metadata=target_metadata,
             include_object=include_object,
-            # Required for migrations that use autocommit_block().
-            transaction_per_migration=True,
-            # Concurrent DDL cannot run inside a transaction owned by the
-            # caller.  Migrations use regular DDL for that compatibility path.
-            knowhere_external_transaction=caller_owned_transaction,
         )
 
         with context.begin_transaction():
@@ -152,7 +136,7 @@ def run_with_connection(connection: Connection) -> None:
         return
 
     if isinstance(configured_connection, Engine):
-        with configured_connection.connect() as connection:
+        with configured_connection.begin() as connection:
             run_with_connection(connection)
         return
 
@@ -165,7 +149,7 @@ def run_with_connection(connection: Connection) -> None:
         connect_args=ssl_connect_args,
     )
 
-    with connectable.connect() as connection:
+    with connectable.begin() as connection:
         run_with_connection(connection)