Unify sweep trees into conf/scaling/ + Python metric presets

[philipnickel]

What

Two readability/maintainability refactors of the Hydra/hydra-zen config surface, verified to leave the composed configs unchanged.

1. conf/scaling/ axis group — single source of truth

The swept sim.* values were duplicated across two parallel trees (conf/sweep/* for timing, conf/nsight_sweep/* for NCU):

value	before	after
9-value particle list	4 files	1 (scaling/particle_count), shared
${ppc_grid:…} derivation	2 files	1 (scaling/weak)
density grid list + N=20M	2 files	1 (scaling/density)
sm operating point	2–3 files	1 (scaling/sm)

Both wrapper trees now pull the axis via defaults: - /scaling: <axis> and keep only their genuine differences (backend list ±jax, plots vs nsight_plot callback, output-dir roots).

2. Metric presets as a Python source of truth — nsight_metric_sets.py

conf/nsight_metrics/{roofline,full}.yaml hand-duplicated the roofline counters. FULL_METRICS is now composed as TIMING + ROOFLINE + extras, so the roofline subset is defined once. Registered as the nsight_metrics config group (package nsight.analyze), imported by profile_nsight.py before Hydra composes. conf/nsight_metrics/ is removed.

Verification

• --cfg job/--cfg hydra for all 9 sweep selections + 3 metric presets is byte-identical before/after, except:
  1. the full preset's metric order (set proven identical: 59 metrics, no dupes — order is irrelevant to NCU and the long-format parquet);
  2. the expected new scaling: entry under hydra.runtime.choices.
• 68 tests pass, ruff clean.
• Runtime-validated on an A100 across all five backends: sim=benchmark, both timing sweeps (${ppc_grid:…} derivation confirmed: 100k→G=28, 200k→G=36), and NCU collection driven by the Python presets (roofline 15/15, full 59/59, zero missing), including an nsight_sweep multirun that produced all five figures.

Follow-up (not in this PR)

P2 consistency pass: rename wrappers to the canonical axis names (weak_scaling→weak, particle_density→density), hoist the duplicated gpu_kind resolver, normalize indentation. Two stale .claude/CLAUDE.md pointers to conf/nsight_metrics/ should also be updated (the edit was auto-blocked as self-modification).

🤖 Generated with Claude Code