coral_anneal: 13.8× per-step speedup via incremental energy tracking

[wock9000]

Summary

One-line algorithmic fix in `coral_anneal.py`: the previous `run_mtm_anneal` called `exact_energy(spins, J)` on every accepted Metropolis move (O(N²) — a full `-½sᵀJs` recompute), then took the delta. Replaced with incremental `e_cur += dE_true` since we already compute `dE_true` exactly in int64 arithmetic on each accept; the result is bit-equivalent without the recompute.

This was FINDINGS entry 6's "Coral-side per-step bookkeeping" bottleneck. The diagnosis (C/Cython rewrite of softmax) was wrong — it was an algorithmic bug in our own code, masquerading as Python/numpy overhead.

Result (same instance + seeds as FINDINGS 6: random_spin_glass N=1024, d=0.4, s=42; 4 seeds × 3200 steps)

variant	ms/step	e_best mean	host-to-Coral ratio
Host (M3 Pro) NumPy	0.34	−14463	1×
Coral over-network (PR #47)	68.78	(J-mismatch)	209×
Coral co-located v1 (FINDINGS 6)	43.33	−14749	133×
Coral co-located v2 (this PR)	3.14	−14749	9.4×

13.8× per-step speedup. Cumulative improvement from the original V3-over-network number: 22×. Quality unchanged.

What's NOT in this PR

• FINDINGS revisions (the entry 5 over-claim about "retiring the experiment" and the new entry for this win) — separate PR so the perf number can be reviewed in isolation.
• Wishart-J quality experiment — separate work; requires rebuilding the TPU model with a Wishart-J baked in.
• Further Coral optimizations (TPU matmul pipelining, batched-replica matmul). The remaining headroom on this arc is ~3× more if pipelined, ~10× more if batched — but the next experiment to run is the quality one, not more optimization.

Test plan

• e_best parity confirmed against FINDINGS 6's data (−14947 min, −14749 mean — exact match)
• Quality unchanged (within seed variance)
• Reproduces from `tools/coral_anneal_bench.py` end-to-end

🤖 Generated with Claude Code