refactor: split infer core into subpackages by concern

- Eliminate core/ directory into cache/, runtime/, network/ subpackages plus flat modules
- Split cache.py (647 lines) into cache/{buffer,strategy,pool}.py by layer
- Add explicit ContiguousStrategy, make AllocationStrategy a real ABC
- Move TaskCacheState to cache/strategy.py, drop string forward references
- Rename api/ to network/, server.py to app.py
- Move sample.py into runtime/ alongside executor and graph
- Simplify TaskCacheManager.__init__ to single pool param
- Expose pool.strategy and pool.req_pool as public properties
- Fix KVCache import in attention_backend.py (TYPE_CHECKING guard)
- Fix steady-state decode reading uninitialized position_ids on first step

tests/extension/test_backend_equivalence.py

@@ -7,18 +7,13 @@ seq_lens with padding mask), and end-to-end scheduler.run_batch.
 import torch
 
 from astrai.extension import ATTN_BACKEND, attn_backend
-from astrai.inference.core.cache import PagePool, TaskCacheManager
-from astrai.inference.core.workspace import InferenceWorkspace
+from astrai.inference.cache import PagePool, TaskCacheManager
+from astrai.inference.workspace import InferenceWorkspace
 from tests.extension.conftest import D, skip_no_kernel
 
 
 def _mk_task_cache(pool: PagePool) -> TaskCacheManager:
-    return TaskCacheManager(
-        strategy=pool._strategy,
-        req_pool=pool._req_pool,
-        max_seq_len=pool.max_seq_len,
-        pool=pool,
-    )
+    return TaskCacheManager(pool)
 
 
 def _ws(pool: PagePool) -> InferenceWorkspace:
@@ -184,7 +179,7 @@ def test_decode_mixed_seq_lens_matches_torch(cuda_model):
 @skip_no_kernel
 def test_run_batch_cuda_matches_torch_greedy(cuda_model):
     """Greedy decode (temperature=0) should produce identical tokens."""
-    from astrai.inference.core.scheduler import InferenceScheduler
+    from astrai.inference.scheduler import InferenceScheduler
     from tests.helpers import FakeTokenizer
 
     model, _ = cuda_model

tests/inference/test_cache.py

@@ -11,7 +11,7 @@ from astrai.inference import (
     TaskCacheManager,
     page_hash,
 )
-from astrai.inference.core.workspace import InferenceWorkspace
+from astrai.inference.workspace import InferenceWorkspace
 
 
 def _ws(pool: PagePool) -> InferenceWorkspace:
@@ -27,12 +27,7 @@ def _ws(pool: PagePool) -> InferenceWorkspace:
 
 
 def _make_task_cache(pool: PagePool) -> TaskCacheManager:
-    return TaskCacheManager(
-        strategy=pool._strategy,
-        req_pool=pool._req_pool,
-        max_seq_len=pool.max_seq_len,
-        pool=pool,
-    )
+    return TaskCacheManager(pool)
 
 
 # ---- page_hash ----
@@ -345,7 +340,7 @@ def test_page_pool_paged_task_alloc():
     assert task_cache.task_alloc("t1", list(range(10)))
     state = task_cache._states["t1"]
     assert len(state.pages) == 10
-    assert pool._req_pool.req_to_token[state.req_idx, 0].item() == state.pages[0]
+    assert pool.req_pool.req_to_token[state.req_idx, 0].item() == state.pages[0]
 
 
 def test_page_pool_paged_task_extend():
@@ -354,7 +349,7 @@ def test_page_pool_paged_task_extend():
     task_cache.task_alloc("t1", list(range(4)))
     assert task_cache.task_extend("t1", 4)
     req_idx = task_cache._states["t1"].req_idx
-    slot = pool._req_pool.req_to_token[req_idx, 4].item()
+    slot = pool.req_pool.req_to_token[req_idx, 4].item()
     assert slot >= 0
 
 
@@ -364,7 +359,7 @@ def test_page_pool_paged_task_free_releases_slots():
     task_cache.task_alloc("t1", list(range(8)))
     task_cache.task_free("t1")
     assert "t1" not in task_cache._states
-    assert len(pool._req_pool.free_slots) == 4
+    assert len(pool.req_pool.free_slots) == 4
 
 
 def test_page_pool_paged_bind_roundtrip():

tests/inference/test_protocol.py

@@ -5,9 +5,9 @@ from unittest.mock import MagicMock
 
 import pytest
 
-from astrai.inference.api.anthropic import AnthropicResponseBuilder
-from astrai.inference.api.openai import OpenAIResponseBuilder
-from astrai.inference.api.protocol import GenContext, StopChecker, StopInfo
+from astrai.inference.network.anthropic import AnthropicResponseBuilder
+from astrai.inference.network.openai import OpenAIResponseBuilder
+from astrai.inference.network.protocol import GenContext, StopChecker, StopInfo
 
 
 def _make_ctx(**kwargs):

tests/inference/test_sample.py

@@ -2,7 +2,7 @@
 
 import torch
 
-from astrai.inference.sample import (
+from astrai.inference.runtime.sample import (
     FrequencyPenaltyStrategy,
     SamplingPipeline,
     TemperatureStrategy,
@@ -268,7 +268,7 @@ def test_sample_return_logprobs_matches_manual_computation():
     logits = torch.randn(2, 30)
     tokens, logprobs = sample(logits, temperature=0.7, top_p=0.95, return_logprobs=True)
     # Recompute with the same pipeline
-    from astrai.inference.sample import (
+    from astrai.inference.runtime.sample import (
         SamplingPipeline,
         TemperatureStrategy,
         TopPStrategy,

tests/inference/test_scheduler.py

@@ -38,8 +38,8 @@ def test_scheduler_concurrent_add_task(mock_model_and_tokenizer):
     """Test concurrent add_task operations."""
     mock_model, mock_tokenizer = mock_model_and_tokenizer
 
-    with patch("astrai.inference.core.scheduler.AutoModel"):
-        with patch("astrai.inference.core.scheduler.AutoTokenizer"):
+    with patch("astrai.inference.scheduler.AutoModel"):
+        with patch("astrai.inference.scheduler.AutoTokenizer"):
             scheduler = InferenceScheduler(
                 model=mock_model,
                 tokenizer=mock_tokenizer,
@@ -77,8 +77,8 @@ def test_scheduler_concurrent_add_remove_task(mock_model_and_tokenizer):
     """Test concurrent add and remove task operations."""
     mock_model, mock_tokenizer = mock_model_and_tokenizer
 
-    with patch("astrai.inference.core.scheduler.AutoModel"):
-        with patch("astrai.inference.core.scheduler.AutoTokenizer"):
+    with patch("astrai.inference.scheduler.AutoModel"):
+        with patch("astrai.inference.scheduler.AutoTokenizer"):
             scheduler = InferenceScheduler(
                 model=mock_model,
                 tokenizer=mock_tokenizer,
@@ -126,8 +126,8 @@ def test_scheduler_concurrent_get_stats(mock_model_and_tokenizer):
     """Test concurrent get_stats operations."""
     mock_model, mock_tokenizer = mock_model_and_tokenizer
 
-    with patch("astrai.inference.core.scheduler.AutoModel"):
-        with patch("astrai.inference.core.scheduler.AutoTokenizer"):
+    with patch("astrai.inference.scheduler.AutoModel"):
+        with patch("astrai.inference.scheduler.AutoTokenizer"):
             scheduler = InferenceScheduler(
                 model=mock_model,
                 tokenizer=mock_tokenizer,

tests/inference/test_tool_parser.py

@@ -2,7 +2,7 @@
 
 import pytest
 
-from astrai.inference.api.tool_parser import (
+from astrai.inference.network.tool_parser import (
     _TOOL_CALL_HEAD_RE,
     BaseToolParser,
     SimpleJsonToolParser,

tests/trainer/test_rollout.py

@@ -3,7 +3,7 @@
 import pytest
 import torch
 
-from astrai.inference.core.scheduler import InferenceScheduler
+from astrai.inference.scheduler import InferenceScheduler
 from astrai.trainer.rollout import (
     BaseRewardModel,
     RawRollout,