- Add strip_fixture_format_from_nodeid() to extract base nodeid.
- Add get_all_fixture_format_names() for format name lookup.
- Used to ensure related fixture formats share cache keys.

- Add T8nOutputCache LRU cache class for storing t8n outputs.
- Add t8n_output_cache field to FillingSession.
- Add xdist_group markers during collection for --dist=loadgroup.
- Use t8n-cache-{hash} prefix to distinguish from user-defined groups.
- Strip cache-specific @t8n-cache-* suffix from nodeids in TestInfo.

- Add cache key helpers to BaseTest (_get_base_nodeid, _get_t8n_cache_key).
- Add _get_filling_session() to access cache from test instances.
- Cache t8n outputs in _generate_block_data() for reuse across formats.
- Skip caching for engine_x and engine_sync variants (different execution).

- Test T8nOutputCache LRU behavior, eviction, and hit/miss tracking.
- Test strip_fixture_format_from_nodeid for various nodeid patterns.
- Test get_all_fixture_format_names ordering and contents.
- Test cache key consistency across fixture format variants.
- Test _strip_xdist_group_suffix preserves non-cache group markers.

Add tests to verify that test items are sorted during collection
to ensure deterministic cache hits. The tests demonstrate:

- Sorting groups related fixture formats by base nodeid.
- Without xdist, items are correctly sorted.
- With xdist, items are NOT sorted (BUG causing high variance).
- Expected vs actual behavior comparison.

The xfail test `test_xdist_sorting_required_for_cache_hits` asserts
the correct behavior (sorting with xdist) and fails until the fix
is applied.

- Add helper methods to TransitionToolCacheStats for serialization.
- Initialize aggregated stats on xdist controller in pytest_configure.
- Send worker stats via workeroutput in fixture teardown.
- Add pytest_testnodedown hook to aggregate stats from workers.
- Update pytest_terminal_summary to display aggregated stats.

- Clear `_cache` in `remove_cache()` to prevent stale data leakage.
- Tests without `transition_tool_cache_key` (e.g., state_test) could
  previously retrieve cached results from prior tests via matching
  `call_counter` subkeys.

Sort test items by (is_slow, base_nodeid, nodeid) to optimize execution:
- Slow tests first (LPT scheduling for xdist load balance).
- Related fixture formats grouped together (cache locality).
- Deterministic order within groups.

If ANY fixture format variant of a test is marked slow, ALL variants
are treated as slow to keep them grouped together for cache hits.

Reuses the base_nodeid cache for xdist marker generation to avoid
redundant strip_fixture_format_from_node calls.

BlockchainEngineXFixture and BlockchainEngineSyncFixture had
can_use_cache=False which was dead code (never checked anywhere).
Replace with transition_tool_cache_key="" which is the actual mechanism
that controls caching — empty string means no caching.

For StateTest specs with --generate-all-formats, the _from_state_test
label suffixes cause alphabetical sort to interleave cacheable and
non-cacheable formats: blockchain_test_engine_from_state_test (cacheable)
→ blockchain_test_engine_x_from_state_test (non-cacheable, clears cache)
→ blockchain_test_from_state_test (cacheable, but cache is gone).

Add has_cache_key to the sort key so cacheable formats cluster together
within each base nodeid group, ensuring the second cacheable format hits
the warm cache before any non-cacheable format clears it.

node_id_for_entropy strips fixture format and fork names from the node
ID before hashing it for deterministic address generation. However, it
did not strip the xdist @group_name suffix (e.g., @t8n-cache-abc12345),
causing different addresses when running with vs without xdist workers.

Strip the suffix so addresses are deterministic regardless of whether
xdist is active.

Replace the raw hit/miss counts with an efficiency metric where 100%
means all tests that could have hit the cache did hit it. Track unique
cache keys to compute expected hits (total cacheable - unique keys).

Also filter subkey stats to only count cacheable tests, eliminating
phantom misses from non-cacheable tests that still interact with the
OutputCache after remove_cache().

Before: T8n cache: key_hits=6, key_misses=6 (50.0%), subkey_hits=6, subkey_misses=18 (25.0%)
After:  T8n cache: 100% hit rate (6/6 expected), 6 t8n calls saved

Pydantic's ModelMetaclass caches __init__ wrappers for dynamically
created classes. When pytester runs multiple fill sessions in-process,
cached wrappers from an earlier session re-invoke __init__ re-entrantly
in later sessions, causing generate() to run twice per test and
doubling the opcode count.

- Switch fill tests to runpytest_subprocess() for process isolation.
- Normal `fill` runs are unaffected (each invocation is a fresh process).

The rebase introduced a second pytest_testnodedown definition for cache
stats aggregation, shadowing the existing timing logs hook. Extract the
cache stats logic into _aggregate_cache_stats() and call it from the
single hook.

Remove dead code that was never called: _get_base_nodeid(),
_get_t8n_cache_key(), _get_filling_session() from BaseTest and
remove_opcode_count() from TransitionTool. Also remove the
now-unused strip_fixture_format_from_node import, TYPE_CHECKING
import, and guarded FillingSession import block from base.py.

Add a clear() method to OutputCache to encapsulate clearing the
internal _cache dict and resetting the key, instead of having
TransitionTool.remove_cache() reach into private members directly.

Also fix the set_cache docstring ("LRU behavior" → "single-key
eviction"), fix the cached_result truthiness check to use
`is not None` for defensive correctness, and fix the node()
docstring to say "pytest node" instead of "node ID".

The @t8n-cache-* xdist group suffix was leaking into the test_id and
group_salt stored in pre-alloc groups, causing fixture output to differ
from runs without xdist. Use _strip_xdist_group_suffix() (already used
by node_to_test_info) on both the group_salt fallback and the test_id
passed to add_test_pre.

Test set_key eviction, get/set round-trip, hit/miss counter
accuracy, clear() behavior, and state across key changes.
Uses sentinel objects as lightweight stand-ins for
TransitionToolOutput.