Implement Add Model Work Label

Own the independent model.work accountant for one architecture family. The
accountant answers what work the model must perform from the model config and
workload alone. Never derive minimum FLOPs or bytes from simulator kernels,
CostTree shapes, profiler rows, or achieved-work logs; doing so would make
redundancy self-referential.

Use the simulator tree only after the semantic work inventory is fixed, to map
semantic rows onto stable, exact location names.

Preconditions

• top-split-model-into-kernels has established the true architecture from HF
  modeling code, a paper, or another authoritative implementation.
• The model config exists and its exact architectures[0] value is known.
• Before writing a location map, the L4 arch is wired and its manifest/location
  names are stable. The independent builder and golden tests may be written
  earlier.

If the architecture math is unresolved, return to top-split-model-into-kernels.
If location names are still changing, finish the L4 wiring before mapping them.

Read First

• model/work/README.md for the accountant, segment, parameter, and validation
  contracts.
• The target model/config/*.json and the architecture evidence produced by
  exploration/splitting.
• The closest files under model/work/attention/, model/work/ffn/, and
  model/work/models/.
• model/work/registry.py and tests/test_model_work.py.
• For location maps, analyzer/rust/src/optimality/location.rs, the final
  simulator manifest/location names, and the closest file under
  model/work/location_maps/.

Workflow

1. Establish the independent inventory

Write down, from the true model implementation:

• exact architecture string and layer schedule;
• attention mechanism, head geometry, cache/state shape, and prefill/decode
  behavior;
• FFN dimensions, routed/shared experts, top-k, and router behavior;
• embeddings, output head, norms, parameter tying, and dtype assumptions;
• formulas for compulsory FLOPs, parameter bytes, and persistent state traffic.

Do not use ServingStudio Sim's partitioned or fused kernel shapes as evidence for these
formulas.

2. Compose or extend the accountant

• Reuse an existing AttentionSpec or FFNSpec when its math matches.
• Add a mechanism under model/work/attention/ or model/work/ffn/ only when
  the math is genuinely new.
• Keep the per-model builder under model/work/models/ thin: parse config fields
  and compose specs.
• Represent heterogeneous schedules with LayerStack; use stable tags that
  produce meaningful semantic segment prefixes.

3. Register the architecture

Register the builder in model/work/registry.py under the exact
architectures[0] string. Preserve the hard failure for unknown architectures;
silent fallback would label unsupported models incorrectly.

4. Add hand-derived goldens

Extend tests/test_model_work.py with independently calculated expectations for
representative prefill and decode workloads:

• total and activated parameter counts;
• FLOP buckets and compulsory byte buckets;
• scaling invariants for sequence length, sampled positions, layer schedules,
  and expert activation where applicable.

Do not calculate expected values by calling the implementation under test.

Run:

uv run pytest tests/test_model_work.py
uv run python -m model.work <config> --prefill 8192@0 --json
uv run python -m model.work <config> --decode 256x4096 --json
uv run python -m model.work.parameter_counts <config>

Choose smaller representative shapes when the architecture requires different
inputs, but cover both prefill and decode semantics.

5. Add semantic location maps

Add one versioned map per supported architecture/deployment location set under
model/work/location_maps/.

• Match exact arch_types and exact non-communication CostTree locations.
• Consume every semantic row exactly once.
• List every non-communication location exactly once.
• Use an empty semantic list when a location has zero compulsory work under the
  cross-leaf-fusion convention.
• Never infer minimum-work formulas from the simulator location being mapped.

If a deployment has no correct map, report it as unsupported instead of claiming
full R6/R7 readiness.

6. Verify the analyzer handoff

Run a small simulation for each mapped deployment and inspect one exact,
batch_locked iteration through the analyzer. Confirm:

• necessary_work_available=true and the expected mapping id is reported;
• every non-communication kernel receives the intended necessary-work mapping;
• aggregate minimum work reconciles with model.work for the same workload;
• R6 necessary-work coverage and R7 redundancy are present and numerically sane.

Use operate-run-simulation to launch the run and operate-use-analyzer to read
the exact iteration; do not substitute sweep-level or UI-derived numbers.

Completion Report

Return the architecture string, reused/new specs, builder and registry paths,
golden-test results, supported location-map/deployment ids, and exact-iteration
R6/R7 evidence. Explicitly list any deployment that remains unmapped.