feat: nonlinear deposit curve (lever 4)

[tamirelazar]

Lever 4 — Nonlinear deposit

Adds a per-frame deposit-accumulation buffer with a selectable nonlinear curve so density spikes compress into filaments instead of blowing out to flat blobs. Fully off by default — byte-identical to current behavior.

How it works

Agents accumulate their existing per-agent unit (deposit_amount · dt) into a per-TrailMap scratch buffer (mirrors the lever-7 blend_src pattern). Once per frame, before diffuse/decay, a fold pass applies curve(accum) · scale, an optional clamp cap, adds the result to the trail, and zeros the buffer.

The accumulation path is gated: when curve=linear && scale=1 && cap=0, the buffer is never touched and deposits go straight to the trail (the existing code path) — so the off case is byte-identical by construction, avoiding the float-reassociation drift a naive accumulate-then-fold would introduce.

Curves

• linear — x (identity, default)
• sqrt — x.sqrt() (compresses density → filaments)
• log — ln(1 + x) (log1p; 0 at 0, guards log(0))
• pow — x.powf(gamma) (--deposit-gamma is the exponent)

Knobs

--deposit-curve <linear|sqrt|log|pow>, --deposit-scale, --deposit-gamma, --deposit-cap (0 = off). All on SimConfig (mirrors lever-7 decay_gamma/diffuse_weight); CLI flags + saved-config round-trip (serde(default) back-compat). Buffer is per-TrailMap, giving per-species structure in separate-trails mode for free. Per-preset tuning is intentionally deferred to the showcase-preset work.

Testing

• Unit tests per curve; deposit_off_path_is_deterministic and a linear-accum fp-equivalence test (active identity config matches the direct path within a magnitude-relative tolerance, accounting for float reassociation); deposit_sqrt_curve_changes_output.
• Saved-config round-trip + old-TOML-loads-with-defaults back-compat.
• cargo test 800 pass · cargo test --all-features --lib 812 pass · cargo clippy --all-targets --all-features -- -D warnings clean.
• 47/47 visual-regression goldens unchanged (defaults are the off path).
• Empirical same-seed capture: linear and pow γ=1 identical to baseline; sqrt/log/pow γ=0.5 differ sensibly.
• Added bench_fold_deposits_sqrt.

Closes #31.