…PLING routing

The condition for LOCAL_WITH_SAMPLING should be based on whether the
node is a narrow transform, not whether it has default settings.
nodes_with_defaults is about whether a node can run without user
config — it's unrelated to execution strategy.

Narrow transforms (select, sample, union) are lightweight column-level
operations safe to compute locally. Wide transforms (sort, record_count)
reshape data and should always go REMOTE.

Updated tests to reflect that sort goes directly to REMOTE from
_determine_strategy, not via the override in execute().

https://claude.ai/code/session_01DaTkkXLAJyC2EfopwiY55d

When cache_results is enabled on a node, the result must be fully
materialized and stored by the remote worker. This means even narrow
transforms (select, sample, union) that would normally use
LOCAL_WITH_SAMPLING must go REMOTE.

Strategy priority in _determine_strategy:
1. local → FULL_LOCAL
2. cache_results → REMOTE
3. narrow transform → LOCAL_WITH_SAMPLING
4. everything else → REMOTE

Added TestCacheResultsStrategy with 5 tests covering cache override
for narrow/wide transforms and the full decision flow.

https://claude.ai/code/session_01DaTkkXLAJyC2EfopwiY55d

Remove verbose debug info from log line ([node_type=..., transform_type=...,
has_default=...]) and the unused initial_strategy variable. The log now
shows just: "Starting to run select (NEVER_RAN -> LOCAL_WITH_SAMPLING)"

Add "Execution Strategy" section to the Core Internals developer docs
covering the three strategies, the decision pipeline, LOCAL_WITH_SAMPLING
vs REMOTE flow diagrams, and node classification table.

https://claude.ai/code/session_01DaTkkXLAJyC2EfopwiY55d

The docs described the strategy routing but didn't connect it to the
execution modes visible in the frontend. Restructured to lead with
what users see (Development vs Performance mode) and then explain the
underlying strategies. Added a section for each mode, clarified that
LOCAL_WITH_SAMPLING only applies in Development mode, and updated the
node classification table with a Performance mode column.

https://claude.ai/code/session_01DaTkkXLAJyC2EfopwiY55d

Sort (reordering) is a wide operation. Narrow transforms do dropping,
renaming, and filtering.

https://claude.ai/code/session_01DaTkkXLAJyC2EfopwiY55d

Tests the full preview data path:
- TestPreviewAfterExecution: preview returns sample data after run,
  caps at 100 rows, returns empty before run, doesn't recompute,
  caches across repeated calls
- TestPreviewAfterUpstreamChange: documents that node_stats.has_completed_last_run
  is NOT reset when settings change, causing stale preview data.
  Also verifies preview is correct after re-running.

Adds create_graph_with_read_and_select helper using fake_data.parquet
(1000 rows) for integration-level preview testing.

https://claude.ai/code/session_01DaTkkXLAJyC2EfopwiY55d

When sibling nodes run in parallel and both serialize the same
upstream LazyFrame, Polars' internal RefCell can raise "Already
borrowed". This is a thread contention issue in the core process
during LazyFrame serialization, not a logic error.

After a parallel stage completes, any node that failed with
"Already borrowed" is now retried sequentially — the contention
is gone because no other thread is accessing the shared LazyFrame.

https://claude.ai/code/session_01DaTkkXLAJyC2EfopwiY55d

When sibling nodes run in parallel and both access the same upstream
LazyFrame, Polars' internal RefCell raises "Already borrowed".

Fix: each node acquires its input nodes' _execution_lock before
calling _function(*input_data), holding it through the LazyFrame
access. Sibling nodes sharing a parent now serialize their access
to the parent's data. Nodes with different parents still run fully
parallel.

Reverts the retry-based approach in favor of preventing the
conflict entirely.

https://claude.ai/code/session_01DaTkkXLAJyC2EfopwiY55d