fix(strategy): size reconstruct() row-block from compute_dtype, not out_dtype

[boskodev790]

Lane

fix / bug — corrects repository behavior. CPU-safe validation only; no GPU scorecard.

Problem

strategy.subspace.reconstruct() streams its (n × n) output row-by-row to stay inside a VRAM budget, sizing the row-block from the output dtype's itemsize:

blk = _row_block(n, n, backend, np.dtype(out_dtype).itemsize, frac)

But every device tensor inside the loop — Q, Ctil, QT, and the (rb, n) product from backend.matmul — lives in cfg.compute_dtype, not the output dtype. For dtype="fp16" those differ: compute_dtype is bumped to fp32 for accuracy (see Config.compute_dtype). So the row-block is sized for half the real per-row footprint, and on an actual GPU at --dtype fp16 with large N / tight vram_fraction the streamed path can OOM. This is the same class of under-budget as the auto_tile operand-upcast fixes, on the reconstruct path.

Fix

Size the block from the dtype the loop actually computes in — mirroring how compress() / stream_gemm_* already pass compute_dtype:

def reconstruct(Ctil, Q, C_out, backend, out_dtype, frac=..., compute_dtype=None):
    item_dtype = compute_dtype if compute_dtype is not None else out_dtype
    blk = _row_block(n, n, backend, np.dtype(item_dtype).itemsize, frac)

and pass cfg.compute_dtype at the call site. compute_dtype=None defaults to out_dtype, so the existing 5-arg call in test_subspace.py is unchanged. Only strategy/subspace.py and tests/ change — no scorer/eval/docs/CI paths.

Tests

Added tests/test_reconstruct_row_block.py (CPU-only, no GPU — pure sizing arithmetic via a fake backend + a _row_block spy):

• fp16 output + fp32 compute → the row-block is sized from fp32 (4 bytes), not fp16 (2) — fails on the pre-fix code, passes after.
• compute_dtype=None falls back to out_dtype (back-compat).
• output math (Q @ Ctil @ Qᵀ) is unchanged.

uv run python -m strategy.smoke
uv run --extra test python -m pytest tests/ strategy/tests/ eval/tests/ -q
# 81 passed, 23 skipped

Fixes #66