Card 78 — Rust kernel proven bit-exact against the oracle S3 DONE · ≤1e-9 GREEN

Last iteration built the answer key; this iteration built the engine part and proved it gives the same answers. The production pipeline computes everything in Rust (a fast systems language); the trusted reference from iteration 01 was Python. Today the Rust version of Hoeffding's Φ² — the "is move size related to bar duration, in any shape?" measurement — was hand-written and then made to reproduce all 19,602 frozen reference answers (9,801 per market, Bitcoin and Ethereum) to within 0.000000001. It passed on the first full run.

Why so paranoid about "the same answers"? This feature is built on rankings (which bar had the 1st, 2nd, 3rd… biggest move). Rankings are brittle: if two moves are nearly identical and two systems round the 16th decimal differently, the ranking flips and the final number jumps visibly. A past feature (Bartels, incident I2) failed exactly this way. The fix, reused here: both systems rank a quantity computed by a single division — an operation the floating-point standard guarantees is bit-identical everywhere — so a ranking flip is impossible by construction.

The kernel was also cross-examined three independent ways: (1) a deliberately slow, dumb counting version written in Rust re-derives every number from scratch and must agree — a bookkeeping bug in the fast version can't confirm itself; (2) a property battery checks known truths (perfectly linked inputs score near 1, the score never leaves its range, scaling all prices by 4 changes nothing — bit for bit); (3) the Python generator already double-checked itself with exact whole-number arithmetic in iteration 01.

One genuine discovery: the textbook says Φ² lives in [0, 1] — true for the idealized continuous formula. Our finite 199-pair grid version can reach 1.0000631 (a hair over 1) but ONLY under mathematically perfect lock-step dependence, which real markets never produce. This was found because a property test honestly failed; it's now precisely documented in three places rather than papered over.

Plumbing that came along: the engine's "memory" (checkpoint) — what lets it resume seamlessly after a restart — now also carries the rolling window of bar durations this feature needs. Old checkpoints from before this change still load fine: the feature simply warms up for 200 bars while every existing feature continues uninterrupted.

The one red light, and why it's good news: a watchdog test ("schema contract") now fails on purpose. It noticed the new measurement exists in the engine but not yet in the database blueprint — the exact class of silent mistake that once shipped 8 features whose values were computed and then thrown away (incident #556). The watchdog stays red until the next iteration wires the database side, which is precisely its job.