SO3onS2: fix selective-map rank deficiency for L >= 8; key CUDA graph on input signature

[johmathe]

Problem

For a real signal each bispectral entry is a single real number (real if l₁+l₂+l is even, imaginary if odd) and
β_{l₁,l₂,l} = (−1)^{l₁} √((2l+1)/(2l₂+1)) · β_{l₁,l,l₂}.
So the "chain" rows (a, b, l) and "cross" rows (a, l, b) that _proved_linear_bootstrap_block used for l ≥ 8 were the same equation. The per-degree block had real rank ≤ l+1 < 2l+1, and the whole selective invariant was rank-deficient for L ≥ 8 (deficit ⌊(l−6)/2⌋ per degree; 20 dims at L = 15). The map also carried identically-zero entries ((l,l,odd) and odd-parity triples with a repeated index).

Fix

• _distinct_pair_linear_block(l): 2l+1 rows (a, b, l) over distinct unordered pairs a < b < l, skipping identically-zero triples.
• Identically-zero entries filtered at every degree; self-coupling fallback restricted to even l ≥ 2 so freed budget is not spent on redundant (l, l, 0).
• Completeness docstrings now state what is verified numerically (real Jacobian rank (L+1)²−3 for L ≤ 16 in tests; per-degree blocks to l = 100 via the benchmark) vs. what is conjectural (global separation). The "proved" wording is gone.
• benchmarks/verify_linear_bootstrap.py rewritten as a per-degree real-Jacobian rank check (seed block l ≤ 3 + degree blocks); passes l ≤ 100, σ_min ≈ 5 at every degree.
• Also: CUDA-graph cache keyed on (shape, dtype, device) instead of batch size (a float64 input no longer replays a float32 graph); sparse CG cache written atomically (tmp + os.replace), EOFError tolerated on load.

Breaking

output_size / layout change for every lmax ≥ 3 (L=3: 15, L=8: 124, L=16: 426 — was 430). The on-disk sparse CG cache self-invalidates. README table updated.

Tests

Closed-form block check; no identically-zero entries; linear rows are distinct equations; whole-invariant Jacobian rank == (L+1)²−3 for L = 4…16 with SVD gap > 1e6 (fails on main for every L ≥ 8); per-degree block rank for l = 4…20; CUDA-graph dtype/shape cache tests (CPU mocks + CUDA parity).

Part of the v0.4.0 stack (D). Note: PR #48 touches the same files and will need a rebase or to be closed.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.59%. Comparing base (b3746db) to head (fb3527a).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #56      +/-   ##
==========================================
+ Coverage   95.53%   95.59%   +0.06%     
==========================================
  Files          11       11              
  Lines        2217     2226       +9     
==========================================
+ Hits         2118     2128      +10     
+ Misses         99       98       -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.