- Generic[E] dispatch: wrapper doesn't resolve, protocol rejection for
  non-conforming executor, inactive function + executor returns None,
  async_call routes through async_execute_callable
- LocalExecutor callable: async fn via execute_callable, sync/async fn
  via async_execute_callable
- CachedFunctionPod: same tag+different packet cached separately, same
  tag+same packet is cache hit, inactive function doesn't store,
  output_schema delegation, dual caching (result_database +
  pod_cache_database)
- Pipeline: no opts uses engine directly (no with_options call)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- CachedFunctionPod now aligns with CachedPacketFunction storage: stores
  function variation data, execution data, cache entry hash, and timestamp
- Cache entry hash computed from tag + system tags + input packet hash
  (matching pipeline record entry_id pattern), ensuring two rows with
  identical user tags but different system tags are cached separately
- FunctionNode.attach_databases() creates CachedFunctionPod wrapping the
  function pod instead of CachedPacketFunction wrapping the packet function
- FunctionNode.process_packet() delegates to CachedFunctionPod for result
  caching and separately records pipeline provenance entries
- CachedFunctionPod.async_process_packet() does sync DB caching + async
  computation via inner pod's async_process_packet
- add_pipeline_record() now explicitly extracts source columns using
  select() instead of rename-then-drop pattern
- iter_packets Phase 2 skip-check uses single cache entry hash
- Added DESIGN_ISSUES.md note (CFP1) about potential optimization of
  reusing entry_hash between CachedFunctionPod and add_pipeline_record

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

CachedFunctionPod now caches by input packet hash only — the function
output depends solely on the packet, not the tag. Tag-level uniqueness
(tag + system tags + input packet hash) is handled by FunctionNode's
pipeline record (add_pipeline_record / compute_pipeline_entry_id).

iter_packets Phase 2 skip check now uses pipeline entry_ids (which
include tag + system tags + packet hash) retrieved from the pipeline
database, ensuring correct deduplication when the same packet appears
with different tags/system_tags.

Also:
- Extracted compute_pipeline_entry_id() as a reusable method
- Updated DESIGN_ISSUES CFP1: shared ResultCache refactor suggestion
- Added TODO notes for match tier support (aligned with P6)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

New test file covering the dual-database caching architecture:

- compute_pipeline_entry_id: determinism, tag/packet sensitivity
- System tag awareness: identical user tags with different system tags
  produce different pipeline entry_ids
- Result DB vs pipeline DB record counts: same packet/different tags →
  1 result record, N pipeline records
- Phase 1/2 with pipeline entry_ids: Phase 1 yields existing, Phase 2
  skips matching entry_ids, processes only novel combinations
- Same packet + new tag triggers Phase 2 (novel entry_id) even though
  CachedFunctionPod has a result cache hit
- Pipeline records reference same result UUID for identical packets
- Pipeline records include source columns but not data columns of input

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- PythonFunctionExecutorProtocol now extends PacketFunctionExecutorProtocol
  (fixes type bound mismatch and ensures supports()/supported_function_type_ids
  are always available)
- PacketFunctionBase._executor typed as E | None (type-safe access to
  execute_callable on PythonPacketFunction without casts)
- LocalExecutor.execute_callable handles nested event loops (mirrors
  PythonPacketFunction._call_async_function_sync pattern)
- Pipeline._apply_execution_engine always calls with_options() per node
  — executor decides what to copy vs share
- Fixed stale docstring (pod_cache_database: "input packet content hash"
  not "tag+packet hash")
- Fixed type annotations: list[str] | None in test helpers,
  list[Any] for mock executor call lists
- Updated pipeline tests to check node's executor (not original mock)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>