feat(adr-001 #3): adaptive Neumann depth from coherence + tolerance

[ruvnet]

Summary

Today callers of `solve_on_change_sublinear` / `solve_single_entry_neumann` pick `max_terms` blindly. The closure size scales with depth, so an over-pick wastes work; an under-pick silently misses the tolerance bound. This PR lands the math-driven helper:

```rust
optimal_neumann_terms(coherence, b_inf_norm, min_diag, tolerance) -> Option
```

For strict-DD `A = D - O` with coherence `c`, the Neumann series satisfies `‖y_k‖ ≤ ρ^k · ‖y_0‖` with `ρ ≤ 1 - c`. To hit tolerance:

```
k ≥ log(‖b‖_∞ / (min_diag · tolerance)) / log(1 / (1 - c))
```

Returned `k` is clamped to `[1, 64]` (caller-protective floor + cap).

Composition

• Cache `(coherence, min_diag)` once at matrix-build time.
• For each event, compute `auto_terms = optimal_neumann_terms(...)` instead of guessing.
• Pass `auto_terms` as `max_terms` to `solve_on_change_sublinear`.

Example update

The event-driven example (PR #33, #35) now picks `max_terms` adaptively from the matrix's coherence at startup, removing the last hand-tuned magic number from the canonical inner loop.

Discovery: the prior hand-picked `max_terms=24` was under-tuned for the test matrix's coherence (0.2). The math-driven helper picks 64 (clamped) for the same tolerance. Per-event latency goes from ~2 ms to ~12 ms — that's correctness paying its real cost. Higher-coherence matrices (≥ 0.5) would pick lower term counts, pulling the bench crossover lower.

Test plan

• 8 new `coherence::optimal_terms_*` tests covering high/low coherence, edge cases, rejection paths
• 31/31 coherence tests pass
• `cargo run --release --example event_driven_anomaly` runs end-to-end
• Full CI

🤖 Generated with claude-flow