Entry 169

The port played against the original

2026-10-1656mtest

The port is now proved against the original by playing both. rugrats_hle::play boots the original into any level as the oracle probes did (the level patched into 0x80018824 for the boot, the global 0x800c5678 set at the loader 0x80017c0c, 0x8004472c branched past the attract wait, play_cine 0x80074958 returning 0 unless cinematics are wanted), about a minute a boot; tests/hle.rs plays the same pad script on the port and on the original and compares the player each game frame: position, heading (+0x24), state (+0x18c), the camera's heading (atan2(-m02, m22) of the matrix at +0x224), and on request the air velocity (+0x170). The tests are ignored by default (a boot a minute): cargo test --release -p rugrats --test hle -- --ignored --test-threads 1.

What the comparisons found, and the port now does as the original:

• Input latency: a press reaches the original's player a game frame later (the pad is read at the vertical blank). The harness holds the port's pad back by that frame; the port itself stays as quick as the host's input.
• The first frame from a standstill only changes the state (0x8005d11c: probes, slide or climb test, state 1 or 2); the step is the walk's or run's handler's, from the next frame. The port stepped at once.
• The step comes before the turn: the handler steps along the heading as it is (0x8005caac), the step is added (0x8005cb48), then the turn (0x8005bc80). The port turned first; from a standstill with a turn it went 2048 -> 1888 -> 1728 a step ahead of the original.
• The step's rounding: 0x8005caac sets the way (+0x178) to the step's length with 0x80076944 (VectorNormal, scaled, each axis toward zero). The port's >> 12 floored negative steps: a unit a frame of drift.
• No steering in the air or while landing: the heading holds from the takeoff to standing again (1246 throughout in the run-and-jump test).
• The air step in integers, in the body's frame: the read of the original's stack (0x800616ac, 0x800616d4) gave the push 12 an axis before any turn (the branch that scales each axis by 340 >> 15, then 0x4a90be59 >> 51 - not the other, which gives 9), turned by the camera (RotY) into the world and by the heading's transpose into the body, sideways halved; the drag is 32/4096 of the speed's length (an isqrt, 0x8001f4ec) along its VectorNormal, both toward zero; the velocity is kept in the body's frame (+0x170) and each step turned back into the world. Worklog 164's fitted 12.55 and the float step are gone; the original's velocities are matched exactly frame by frame.
• The launch is built along the heading at the takeoff, not at the press (the original's +0x170 (-1, -1000, 679) for a run at 17 facing 1246).
• The bump's end: state 9 reads no stick, so the original turns a frame after the bump ends; the port turned on that frame.

The hub run along the hall into the wall now matches the original exactly on every compared field, every frame; the run and jump from the hub's start matches exactly through the takeoff and within 1 to 4 in the air. Tests that had been aligned by hand to the oracle's first moving frame are shifted by the standing frame.