Refactor architecture domains and add evidence-backed cost modeling

[reiase]

Motivation

Blueprinting is an evidence-driven hardware architecture exploration system. The previous source layout still mixed workload semantics, logical mapping, formal derivation, hardware descriptions, cost evidence, and external validation under compiler-shaped packages. That made ownership unclear, encouraged target details to leak into portable plans, and made the base package carry a large optional performance-data bundle.

This PR converges the implementation around the product architecture: formal modeling and verified synthesis are implementation techniques, while workload, mapping, system, evidence, application, and validation remain explicit domains.

What changed

Establish explicit domain boundaries

• Rename the formal derivation package from blueprinting.compiler to blueprinting.synthesizer.
• Promote evidence and cost analysis to top-level blueprinting.analysis.
• Introduce:
  • blueprinting.schema for dependency-free canonical serialization and immutable values;
  • blueprinting.workload for model and scenario facts;
  • blueprinting.mapping for TP/PP/DP strategies and explicit deployment mappings;
  • blueprinting.system for compute, memory, interconnect, and system profiles;
  • blueprinting.validation for Calculon/Vidur adapters and regression gates.
• Add AST-based dependency-boundary tests to keep the supported architecture acyclic.

Clarify exact-work derivation and late binding

• Move Transformer training and inference work algebra into sibling synthesizer dialects.
• Separate training workload facts from logical mapping choices.
• Keep physical network-tier selection in NetworkTierBinding, supplied only during target/system evaluation.
• Remove duplicate invocation objects from PortablePlanIR; cost views reconstruct semantic metadata from canonical WorkloadFacts.
• Reject conflicting canonical and legacy configuration aliases before derivation.
• Bump changed Transformer pass contracts and planner revisions to v2.

Add the performance evidence plane

• Add a typed cost query/result protocol, resolver, database, provenance records, and explicit fallback policy.
• Add analytical roofline, simulator/tabular import, AIConfigurator import, and Vidur profile support.
• Keep estimates as derived views instead of writing latency into workload or portable-plan facts.
• Move the optional AIConfigurator snapshot to data/evidence/ and exclude it from base distribution artifacts.

Strengthen application, validation, and packaging contracts

• Add presentation-neutral shared application reports for training and inference.
• Replace the legacy root CLI with a small argparse entry point and lazy NiceGUI launch.
• Keep Calculon and Vidur downstream as comparison oracles; consolidate historical SeqSel reproductions under validation/legacy/.
• Add strict training and inference baseline gates, late-binding invariants, canonical/legacy boundary tests, CLI tests, and wheel-content checks.
• Package model/system presets for installed workbench use while excluding optional evidence.
• Update the bilingual MkDocs design narrative, ADRs, status pages, pass documentation, and architecture SVGs.

Architectural invariants guarded by this PR

• Workload contracts do not depend on mapping or hardware.
• Target/deployment details do not enter ModelIR, DistributedTaskIR, or PortablePlanIR.
• Changing a bound network tier can change a cost projection but cannot change the portable-plan digest.
• Exact operations, bytes, shapes, and collective payloads have one canonical source of truth.
• External baselines cannot influence lowering or cost-model construction.
• Legacy wire tags remain readable even though the Python package layout changes.

Compatibility and migration notes

• The blueprinting.compiler Python import path is intentionally removed.
• Historical compiler.* codec tags remain as stable wire identities.
• Retained Calculon-style configuration keys are accepted only at import boundaries; canonical and legacy spellings must agree when both are supplied.
• The package root now exposes version metadata instead of re-exporting the legacy simulation stack.
• AIConfigurator evidence must be loaded explicitly from data/evidence/; it is no longer included in the default wheel or sdist.

Validation

• uv run pytest: 131 passed, no warnings.
• Ruff lint and format checks pass across the supported architecture.
• Calculon/SeqSel training and Vidur inference baseline gates pass.
• Documentation i18n check validates 31 bilingual page pairs.
• mkdocs build --strict and SVG XML validation pass.
• Wheel contract passes: 191 files, 1,182,984 bytes uncompressed, required presets present, optional evidence absent.

Suggested review order

1. docs/design/modules.{en,zh}.md and the new ADRs for intent.
2. src/blueprinting/{workload,mapping,system}/ for domain ownership.
3. src/blueprinting/synthesizer/dialects/transformer/ and lowerings for canonical work derivation.
4. src/blueprinting/analysis/cost/ for evidence resolution.
5. src/blueprinting/validation/, boundary tests, and CI/package checks.

Out of scope

This PR does not claim that target plugins, portable-to-concrete planning, resource scheduling, event-level simulation, profiler ingestion, MachineIR emission, or Pareto search are implemented. It establishes and guards the boundaries those capabilities will extend.