feat(mvp-multinode): AWS dollar cost-model simulator for the run_demo arms

[zzylol]

Summary

Standalone Python package (deploy/mvp-multinode/cost_model/) that models the
per-component × per-resource cost of the six run_demo.sh arms and prices
them in USD/month via an editable AWS price book.

• Resource model (model.py): per component (edge agent, data-plane,
  gorilla-merger, MinIO, Thanos query/store/compact, controller; raw_agent,
  serf_gateway, VM) × per resource (CPU vCPU, memory GB, egress GB/mo, EBS GB,
  S3 storage GB, S3 PUT/GET counts), as functions of workload + ASAP config.
• AWS price book (pricing.py, single editable module): us-east-1 on-demand
  — EC2 decomposed to $0.0264/vCPU-hr + $0.0054/GB-RAM-hr (m6i.large reference),
  S3 $0.023/GB-mo + $0.005/1k PUT + $0.0004/1k GET, EBS gp3 $0.08/GB-mo, egress
  $0.09/GB internet / $0.01/GB cross-AZ.
• Calibration (coefficients.py): every constant tagged measured/derived/
  design/assumed with a citation (FINDINGS wire sweep; the 0510 container-summary
  CPU/mem CSVs; holistic-compression intchunk 2.33× ratio + perf(asapedgeprocessor): accumulate intchunk cold parts per block_duration #442; processor
  config.go defaults). Assumed-pending-live coefficients carry # TODO(live).
• CLI (simulator.py): arm comparison + per-arm breakdown + cold-format and
  sampling-p config sweeps. validate.py reproduces the FINDINGS bandwidth
  at 0.00% error.

Headline (MVP workload, us-east-1, 30-day retention)

arm	wire Mbps	total $/mo	dominant term
b0	18.59	131.9	egress
b1	1.50	76.6	storage
b2	3.03	81.5	storage
b3	9.10	101.9	storage
asap	0.98	67.4	compute
asap-gzip	0.14	64.7	compute

ASAP is ~2× cheaper than raw b0 and beats every baseline: it spends more on edge
EC2 CPU (aggregation) but saves big on egress + S3/EBS storage. The cold-format
sweep shows intchunk+batched (#442) ~190k PUTs/mo ($1) vs the pre-#442 tiny-part
path ~11M PUTs/mo ($57) — the ~60× request saving in dollars.

Why standalone vs the control_plane optimizer cost model

Reuses the tco.rs price-book idea + mod.rs sketch-state sizes, but the Rust
optimizer cost is a planner-internal heuristic (abstract dollars, per-query
candidate ranking) and tco.rs is a 2-way before/after Grafana-Cloud TCO.
Neither models the deployed component set, the six arms, the cold PUT-count
economics, or calibrates to FINDINGS. This is an eval artifact next to the
run_demo harness, calibrated to the same runs.

Test plan

• python3 -m cost_model.validate — reproduces FINDINGS Mbps at 0.00% error (exit 0)
• python3 -m cost_model.simulator — prints the MVP arm comparison in $ + per-component breakdown
• python3 -m cost_model.simulator --sweep cold / --sweep sampling
• python3 -m cost_model.test_cost_model — 10/10 invariant tests pass
• stdlib-only (tabulate/pytest optional, fallbacks verified absent on box)

🤖 Generated with Claude Code