fix(storage): thread seed through iota fill so eval couples differ (#104)

[RealDiligent]

Summary

Fixes issue #104: _fill_iota() ignored the seed argument, so generate(..., fill="iota") always produced the same symmetric (i+j)%97 pattern. The eval harness offsets seeds to build distinct A_i and B_i couples (seed+2i / seed+2i+1), but with iota those offsets were no-ops — every couple had A == B and all --pairs were byte-identical, silently benchmarking A @ A instead of general A @ B.

Threads seed into _fill_iota as (i+j+seed) % 97 in both strategy/storage.py and matmul/storage.py, mirroring _fill_random / _fill_lowrank. Reopens the fix from closed PR #105 on current main.

Result

metric	value
accuracy	1.0 (exact baseline unchanged; restores distinct A/B couples for --fill iota)
time complexity	O(N³) unchanged
latency	N/A
VRAM usage	N/A

Regime measured: CPU-only regression tests (tests/test_iota_fill.py). No strategy cost claim.

Raw scorecard (paste python -m eval … output or --json)

Engine/input fix — not a strategy improvement PR.
Local CI (mirrors eval-policy.yml):
  tests/test_iota_fill.py: 5 passed
  pytest tests/ strategy/tests/ eval/tests/: 83 passed, 15 skipped (no GPU)
New tests verify: distinct seeds -> distinct matrices; eval couple offsets -> A != B; couples differ.
Closes #104

Checklist

• I ran the scorer on unseen couples — no hardcoding of seeds/matrices. (N/A: input-generation fix)
• Accuracy and latency come from the same run at the same dtype. (N/A)
• This is an improvement (every cost axis down, accuracy held) or I
  state honestly which axis it trades — see the one rule in CONTRIBUTING.md.
  This PR restores eval harness semantics for --fill iota; not a strategy submission.
• Correctness gates pass:
  python eval/tests/test_eval.py,
  python strategy/tests/test_subspace.py,
  python tests/test_correctness.py.
• I named the device and dtype so a reviewer can reproduce the numbers.

<----ia--------e------- relabel refresh: 2026-07-07T14:23:28Z -->

[github-actions]

Gate chain passed. This PR is queued for the next batched GPU evaluation window.