Add static inference planning and Vidur drift gates

[reiase]

What changed

• [KNOWN] Adds a phase-neutral decoder inference frontend with separate model, mapping, and request inputs.
• [KNOWN] Lowers prefill and decode phase points through ModelIR -> DistributedTaskIR -> PortablePlanIR, with explicit KV state, workspace capacity, workload facts, collectives, implementation requirements, and pass checkpoints.
• [KNOWN] Adds analytical inference costing plus a separate InferenceCostProvider.resolve() contract; external baselines expose only InferenceBaseline.lookup() and cannot participate in lowering or costing.
• [KNOWN] Adds exact-key Vidur raw component-profile alignment, a pinned MIT-licensed Phi-2/A100/TP1 fixture, combined Calculon/Vidur regression contracts, and pull-request CI.
• [KNOWN] Adds bilingual inference/experiment documentation and an SVG architecture diagram.

Why

[INFERRED] This establishes the first auditable inference-planning slice while preserving Blueprinting's planning/evidence boundaries. It deliberately does not present static decoder-block composition as a serving-system simulator.

Review fixes included

• [KNOWN] Costing now reads PlanTask.workload and explicit plan buffers instead of a hidden InferenceInvocation copy.
• [KNOWN] Application outputs are named prefill/decode-step/model-execution time, not TTFT/TPOT/E2E; memory is reported as analytical fit rather than general serving feasibility.
• [KNOWN] Request workload is separate from execution mapping in InferenceAnalysisDraft.
• [COMPUTED] Vidur reports now include non-cancelling component MAPE and maximum component error. The current three-case comparable-subtotal MAPE is 23.58%, while component MAPE is 54.08% and maximum component error is 99.50%; the gate is explicitly classified as drift detection, not accuracy certification.
• [KNOWN] Exact CSV integer keys reject fractional, NaN, and infinite values instead of silently truncating them.
• [KNOWN] A portable workspace buffer is explicit and verifier-supported.

Validation

• [COMPUTED] uv run pytest: 89 passed.
• [COMPUTED] Pinned training/inference regression tests: passed.
• [COMPUTED] Ruff over every staged Python file: passed.
• [COMPUTED] Bilingual source check: 28 page pairs passed.
• [COMPUTED] uv run mkdocs build --strict: passed.
• [KNOWN] Full-repository Ruff is not enabled by this PR because the existing legacy tree currently has unrelated lint debt.

Known limits and follow-ups

• [KNOWN] The implemented template is dense MHA plus non-gated MLP; full norm/residual topology, GQA/MQA, gated MLP, embedding, LM head, sampler, and MoE are not represented yet.
• [KNOWN] Current DistributedTaskIR materializes a local TP block; PP and replica composition remain analytical outside that graph.
• [KNOWN] Decode contexts are recompiled individually rather than specialized from a parametric plan.
• [KNOWN] Arrivals, continuous batching, scheduler/KV allocation policy, contention, tail latency, and end-to-end Vidur simulation remain future discrete-event work.
• [KNOWN] Raw Vidur component-key alignment does not establish full Phi-2 topology equivalence.