Enhance cost management and refactor architecture components

.github/workflows/quality.yml

@@ -58,5 +58,49 @@ jobs:
       - name: Install locked dependencies
         run: uv sync --locked --extra legacy-ui
 
+      - name: Lint supported architecture boundary
+        run: >-
+          uv run ruff check
+          src/blueprinting/__init__.py
+          src/blueprinting/__main__.py
+          src/blueprinting/cli
+          src/blueprinting/schema
+          src/blueprinting/workload
+          src/blueprinting/mapping
+          src/blueprinting/system
+          src/blueprinting/synthesizer
+          src/blueprinting/analysis
+          src/blueprinting/validation
+          src/blueprinting/application
+          src/blueprinting/workbench
+          tests/synthesizer
+          tests/analysis
+          tests/application
+          tests/regression
+          tests/workbench
+
       - name: Run tests
         run: uv run pytest
+
+  package-contract:
+    name: Base wheel contract
+    runs-on: ubuntu-latest
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
+        with:
+          enable-cache: true
+
+      - name: Install Python
+        run: uv python install 3.12
+
+      - name: Build wheel
+        run: uv build --wheel
+
+      - name: Verify wheel contents and size
+        run: uv run python scripts/check_wheel_contract.py dist/*.whl

AGENTS.md

@@ -94,7 +94,7 @@ Pass pipeline 依据声明式 contract 编排，不得依赖具体 Python pass 
 
 Binding 分为独立维度：workload、strategy、target、deployment、calibration。
 
-- workload/strategy 可以逐步特化，但必须显式记录在 typed derivation context；当前代码名为 `CompilationSession`；
+- workload/strategy 可以逐步特化，但必须显式记录在 typed derivation context；当前代码名为 `SynthesisSession`；
 - target/deployment 只能在 portable plan 之后进入；
 - 同一个 `PortablePlanIR` 必须能绑定到多个实质不同的硬件目标；
 - target 变化不得改变 `ModelIR`、`DistributedTaskIR` 或 `PortablePlanIR` digest；
@@ -142,12 +142,18 @@ handling。若 LPU 的 issue cycle/slot 具有 correctness 含义，它在 targe
 
 ## 9. 当前实现边界
 
-当前形式化分析实现仍位于历史 package path `src/blueprinting/compiler/`。该路径为兼容性保留，不定义产品架构。已经实现：
+无领域依赖的 codec 与 immutable schema primitive 位于 `src/blueprinting/schema/`；target-neutral workload contract
+位于 `src/blueprinting/workload/`，逻辑策略与显式 deployment mapping 位于 `src/blueprinting/mapping/`，芯片、memory、
+interconnect 与 system profile 位于 `src/blueprinting/system/`；canonical 表示与形式化推导机制位于
+`src/blueprinting/synthesizer/`，分析与证据评估位于同级 `src/blueprinting/analysis/`，外部 baseline 与回归 gate
+位于 `src/blueprinting/validation/`。Synthesizer 表示 formal plan synthesis 的实现边界，不是产品身份、RTL 综合器
+或独立 Compiler 组件。已经实现：
 
 - 五层 canonical IR 的 immutable schema、serialization 和 structural verifier；其中后两层仍是 experimental contract；
 - stable ID、lineage、typed scalar expression、binding/session；
 - pass contract、analysis cache/invalidation 和 derivation checkpoint；
-- decoder-only Transformer training frontend；
+- decoder-only Transformer workload contract 与 training/inference frontend adapter；
+- compute、memory、interconnect 和聚合 `SystemProfile` contract；
 - `ModelIR -> DistributedTaskIR -> PortablePlanIR` 的 TP、recompute、workload 与 buffer derivation；
 - peak-only / system-evidence cost view 和 Calculon/SeqSel 校准实验。
 
@@ -162,7 +168,13 @@ handling。若 LPU 的 issue cycle/slot 具有 correctness 含义，它在 targe
 
 ## 10. 代码与仓库规则
 
-- 新形式化表示、推导与分析代码在 package 重命名 ADR 通过前进入 `src/blueprinting/compiler/` 对应边界；不得新建平行表示栈。
+- 无领域依赖的 canonical codec、frozen value 与 schema error 进入 `src/blueprinting/schema/`；workload semantic/request contract 进入 `src/blueprinting/workload/`；逻辑 strategy 与 deployment mapping 进入
+  `src/blueprinting/mapping/`；芯片、memory、interconnect 与 system contract 进入 `src/blueprinting/system/`；workload-to-IR adapter、canonical 表示与推导进入
+  `src/blueprinting/synthesizer/`；cost/evidence analysis 进入 `src/blueprinting/analysis/`。不得新建平行表示栈。
+- `SystemProfile` 是当前有限的 compute/memory/network evidence-bearing adapter，不得被描述成已经实现的完整
+  `ArchitectureBlueprint`；`src/blueprinting/types/system/` 只服务 legacy calculator，新代码不得依赖它。
+- `blueprinting.compiler` Python path 已硬切删除；历史 `compiler.*` canonical codec tag 作为 wire identity 保留，
+  未经迁移 ADR 不得改写。
 - IR 对象默认 frozen；语义字段使用 typed dataclass/enum/ID，不使用自由字典代替 contract。
 - 所有公共 derivation/transformation 和 verifier 必须有 positive、negative、round-trip 与 lineage 测试。
 - Python 最低版本为 3.10；不得使用只在更高版本解析的语法，除非先更新 packaging contract。

README.md

@@ -49,11 +49,11 @@ provides:
 - immutable schemas and structural verifiers for the current five-layer IR backbone; only the first three layers have a production derivation slice, while `ConcretePlanIR` and `MachineIR` remain experimental contracts;
 - stable IDs, lineage, schema-versioned serialization, content digests, and typed binding sessions;
 - declarative transformation contracts with analysis invalidation and derivation checkpoints;
-- a typed decoder-only Transformer training frontend;
-- `ModelIR -> DistributedTaskIR -> PortablePlanIR` staged derivation with explicit TP collectives, recomputation, workload,
-  and buffer facts;
-- peak-only and hardware-evidence cost views;
-- a reproducible Calculon/SeqSel calibration experiment.
+- typed decoder-only Transformer training and static prefill/decode frontends;
+- `ModelIR -> DistributedTaskIR -> PortablePlanIR` staged derivation with explicit TP collectives, recomputation,
+  KV state, workload, and buffer facts;
+- peak-only, hardware-evidence, database, and explicit roofline fallback cost views;
+- reproducible Calculon/SeqSel and Vidur comparison gates that remain downstream of derivation.
 
 First-class architecture blueprints, hardware design variables, concrete resource simulation, network/hardware
 simulator adapters, bottleneck/sensitivity reports, energy/area/cost models, and Pareto search are planned product
@@ -80,38 +80,36 @@ pip install -e ".[dev,docs]"
 ## Build the current Transformer workload blueprint
 
 ```python
-from blueprinting.compiler.lowering import (
+from blueprinting.synthesizer.lowering import (
     DistributeTransformerTrainingPass,
     PlanTransformerTrainingPass,
 )
-from blueprinting.compiler.models import (
-    TransformerExecutionSpec,
-    TransformerModelSpec,
-    build_transformer_model_ir,
-    compilation_session_for,
-)
-from blueprinting.compiler.passes import PassManager, PassPipeline
+from blueprinting.synthesizer.frontend import build_transformer_model_ir, synthesis_session_for
+from blueprinting.synthesizer.passes import PassManager, PassPipeline
+from blueprinting.mapping import TransformerTrainingMappingSpec
+from blueprinting.workload import TransformerModelSpec, TransformerTrainingWorkloadSpec
 
 model = TransformerModelSpec.from_mapping("gpt3-175B", model_config)
-execution = TransformerExecutionSpec.from_mapping(execution_config)
-source = build_transformer_model_ir(model)
+workload = TransformerTrainingWorkloadSpec.from_mapping(execution_config)
+mapping = TransformerTrainingMappingSpec.from_mapping(execution_config)
+source = build_transformer_model_ir(model, datatype=workload.datatype)
 
 result = PassManager().run(
     PassPipeline.of(
         DistributeTransformerTrainingPass(),
         PlanTransformerTrainingPass(),
     ),
     source,
-    session=compilation_session_for(model, execution),
+    session=synthesis_session_for(model, workload, mapping),
 )
 
 portable_plan = result.ir
 for checkpoint in result.checkpoints:
     print(checkpoint.pass_name, checkpoint.ir.digest)
 ```
 
-The mapping inputs use the model and execution schemas in `data/`. Invalid topology such as
-`world_size != tp * pp * dp` is rejected at the typed frontend boundary.
+The adapter reads the retained model/execution JSON presets in `data/`, then separates workload facts from the
+logical mapping. Invalid topology such as `world_size != tp * pp * dp` is rejected at the typed frontend boundary.
 
 ## Reproduce the Calculon calibration
 
@@ -152,24 +150,24 @@ Calculon remains an adjacent calibration utility and does not participate in the
 ## Repository layout
 
 ```text
-src/blueprinting/compiler/
-├── ir/              # five canonical IR contracts
-├── models/          # typed semantic frontends
-├── lowering/        # staged derivation passes
-├── analysis/        # exact workload and derived cost analyses
-├── experiments/     # reproducible validation experiments
-├── passes/          # transformation contracts and manager
-└── session.py       # explicit bindings and typed derivation context
-
-src/blueprinting/application/  # framework-neutral analysis service
+src/blueprinting/schema/       # dependency-free codec and immutable schema primitives
+src/blueprinting/workload/     # target-neutral model and scenario facts
+src/blueprinting/mapping/      # logical strategies and explicit deployment mappings
+src/blueprinting/system/       # chip, memory, interconnect, and system profiles
+src/blueprinting/synthesizer/  # canonical IR, exact-work dialects, and verified derivation
+src/blueprinting/analysis/     # evidence protocols, cost resolution, and projections
+src/blueprinting/application/  # framework-neutral analysis services and reports
+src/blueprinting/validation/   # external baselines and strict regression gates
 src/blueprinting/workbench/    # NiceGUI workbench and legacy presentation adapters
 
-tests/compiler/      # current formal-representation and calibration tests
-docs/                # bilingual MkDocs design, reference, experiment, and project documentation
+data/evidence/                 # optional external evidence, excluded from the base package
+tests/                         # domain, derivation, application, and regression contracts
+docs/                          # bilingual MkDocs design, experiment, and project documentation
 ```
 
-The `compiler` package path and names such as `CompilationSession` are current implementation identifiers retained
-for compatibility; they do not define the product architecture.
+`workload`, `mapping`, and `system` own separate input concerns. `synthesizer` derives canonical plans from workload
+and logical-strategy contracts without reading a physical system; `analysis` later evaluates those plans against an
+explicit system, deployment mapping, and evidence snapshot. External oracles remain downstream in `validation`.
 
 ## Development
 

data/evidence/aiconfigurator/README.md

@@ -0,0 +1,14 @@
+# AIConfigurator evidence bundle
+
+This directory is an external performance-data snapshot, not a Python package
+and not a Blueprinting system contract.  Consumers must load individual files
+through an explicit evidence importer and record the selected file digest,
+runtime/backend revision, hardware identity, and measurement protocol.
+
+The imported snapshot does not currently include one repository-level source
+revision manifest.  Treat it as exploratory evidence; do not use it as a frozen
+regression oracle until provenance and license metadata are pinned for the
+whole bundle.  The files retain their original SPDX headers where supplied.
+
+The directory is intentionally excluded from default wheel and source
+distribution artifacts because it is large and optional.