refactor(benchmarks): explicit QUICK_SIZES/LONG_SIZES per spec (+ fix 14-min memory CI)

[FBumann]

TODO (human): one line on why.

Note

Generated by AI (Claude).

What

Follow-up to #31. Two goals:

1. Fix the 14-min CodSpeed memory job. Benchmark selection rework + time/memory CLI unification #31's quick-subset rework ran 3 sizes per model under --quick, including the heaviest builds (basic n=1600 ≈ 550 MiB, knapsack 1M, qp 20000) under memray. Per-PR --quick now runs a lean 2 sizes per model; patterns keep (0, 50, 100).
2. Make the size tiers explicit per spec (your call) — replace the derived quick subset and the long_threshold number with constants visible at the top of each model file:

SIZES       = (10, 50, 100, 250, 500, 1000, 1600)  # --long
QUICK_SIZES = (10, 250)                            # --quick (per-PR)
LONG_SIZES  = (1000, 1600)                          # --long only

Tiers: --quick → quick_sizes; default → sizes minus long_sizes; --long → all. LONG_SIZES = the sizes above each old long_threshold, so default/--long behave exactly as before — only --quick got leaner. Patterns expose the same shape via a QUICK_SEVERITIES = (0, 50, 100) constant and long_sizes=() (every severity runs by default; severity cost peaks at 100, which stays in quick+default).

skip_reason (the shared pytest↔memray gate), introspect, and CLI help updated; long_threshold removed.

Verify

• ruff + mypy clean; harness tests pass (incl. memory↔pytest id alignment).
• Tier check per spec confirms quick ⊆ default ⊆ long and that default/long match the old thresholds.

[codspeed-hq]

Merging this PR will not alter performance

✅ 79 untouched benchmarks
🆕 59 new benchmarks
⏩ 593 skipped benchmarks¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
🆕	Memory	test_build[masked-n=100]	N/A	735.4 KB	N/A
🆕	Memory	test_build[expression_arithmetic-n=250]	N/A	19.6 MB	N/A
🆕	Memory	test_to_solver[gurobi-basic-n=250]	N/A	69.1 MB	N/A
🆕	Memory	test_to_solver[gurobi-milp-n=50]	N/A	1 MB	N/A
🆕	Memory	test_to_lp[cumsum-severity=100]	N/A	29.3 MB	N/A
🆕	Memory	test_to_solver[highs-storage-n=250]	N/A	37.5 MB	N/A
🆕	Memory	test_to_solver[gurobi-sparse_network-n=250]	N/A	48.1 MB	N/A
🆕	Memory	test_to_solver[highs-knapsack-n=10000]	N/A	2.5 MB	N/A
🆕	Memory	test_to_solver[highs-milp-n=50]	N/A	802.1 KB	N/A
🆕	Memory	test_to_solver[gurobi-sos-n=1000]	N/A	3.1 MB	N/A
🆕	Memory	test_build[nodal_balance-severity=100]	N/A	31.5 MB	N/A
🆕	Memory	test_to_solver[gurobi-knapsack-n=10000]	N/A	2.8 MB	N/A
🆕	Memory	test_to_solver[highs-piecewise-n=1000]	N/A	4.2 MB	N/A
🆕	Memory	test_to_solver[highs-basic-n=250]	N/A	55.1 MB	N/A
🆕	Memory	test_to_solver[highs-masked-n=100]	N/A	659 KB	N/A
🆕	Memory	test_to_lp[qp-n=1000]	N/A	20.4 KB	N/A
🆕	Memory	test_build[storage-n=250]	N/A	11.3 MB	N/A
🆕	Memory	test_to_solver[gurobi-nodal_balance-severity=100]	N/A	25.1 MB	N/A
🆕	Memory	test_build[knapsack-n=10000]	N/A	792.3 KB	N/A
🆕	Memory	test_to_solver[highs-merge_balance-severity=100]	N/A	24.9 MB	N/A
...	...	...	...	...	...

ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.

---

_{Comparing perf/quick-ci-time (5c99d67) with master (4db3c76)}

Footnotes

1. 593 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩