Card 60 started — the time-arrow detector; two traps sprung before any Rust exists S1 DONE

Feature 3 of 4 began: a detector for the arrow of time in price data. Play a video of truly random noise backwards and nobody can tell. Play real markets backwards and subtle things look wrong — crashes are sharper than recoveries, quiet grinds differ from panicked unwinds. Card 60 measures one precise slice of that: turn the last 200 closes into a network (bars are nodes; two bars "see" each other if nothing between them closes as high — the horizontal visibility graph this family is named after), then ask each bar: how tightly clustered are the neighbours you can see BEHIND you versus AHEAD of you? The feature is the average gap. Zero = time-symmetric; on real data it sits slightly positive (~+0.055) — the past really does look different from the future.

An honest renaming ships with it. The research card called it "_kld" (Kullback–Leibler divergence), but a true KL divergence needs a density estimator with tuning knobs this project's parameterless rule forbids. What was actually evaluated — and what ships — is a signed mean gap. The column is named …_meangap, matching what the number IS.

The paranoid answer-key process earned its keep twice today, before a single line of engine code:

• Trap 1 — the tie bug. The first draft used a textbook "monotonic stack" to build the graph fast. The reference library (the exact one the research evaluation used) was wired in as a mandatory cross-check on every window — and instantly vetoed the draft: 3 phantom connections on the very first window. Root cause: when several bars close at exactly the same price (common — prices live on a discrete grid), the stack version lets a bar "see" through an equal-height wall that should block it. The fix is a simpler scan that applies the visibility rule literally. Without the cross-check, those phantom edges would have poisoned every downstream number while looking completely plausible.

• Trap 2 — perfection that floating point can't keep. Mathematically, reversing time flips this feature's sign exactly, so the generator asserts that on every window. It failed — by one part in 10¹⁶. Reversing the window changes the ORDER numbers are added in, and computer arithmetic rounds differently for different orders. The check now allows a hair's width (10⁻¹²) and documents why — a lesson that transfers straight into how the Rust tests must be written.

The frozen answers: 19,602 windows across Bitcoin and Ethereum, every one passing four independent checks (fast scan ≡ reference library ≡ brute force on samples, plus the reversal flip).