tdkpin/tdkpin-rs/PARITY_AUDIT.md

Rust player-experience parity audit

Audit date: 2026-08-23

Decision

The Rust rewrite passes the reconstructed-binary requirements for the same game and player experience at the 640x460 logical gameplay boundary. No known raw, placeholder, guessed, partial, or unverified gameplay unit remains in the Rust path. This is a semantic and presentation result, not a claim that a native Rust executable is a byte-identical Win16 implementation.

The authoritative source is the preserved executable with SHA-256 a9022f1894e3e6e21fc42e8f6c932f7c549ca77f63aaa0c488bb9d55d9d0174c, the complete readable C under ../original/reconstructed/, and the zero-gap C ledger in ../original/C_RECONSTRUCTION_FINAL_AUDIT.md.

Requirement matrix

Requirement	Binary/C basis	Rust evidence	Result
Original presentation	All 63 resources round-trip; DAT/bitmap bounds and GDI call sites are reconstructed	Original images/samples are embedded; loading, attract, game, panel, target, effect-seven, high-score, and name-entry scenarios export 640x460 frames	Verified
Collision inventory	All 175 initialized records are typed in OBJECTS.tsv; raw scan order is address-linked	109 type-2 lines, 40 type-1 circles, all type-3/4 mechanisms, layer masks, response fields, and initial activity are represented; relative chains 54-59 and 69-74 have record-exact tests	Verified
Fixed-point/Real48 physics	1000:c79c and Borland runtime arithmetic are reconstructed; the DOS reference probe is bit-exact	10 ms millipixel steps, both speed clamps, four board bounds, first-record selection, signed dot-product tangent response, spin, auxiliary kick, and dynamic-ball transfer use exact-width arithmetic	Verified
Live collision behavior	Stopped Wine probes expose position, velocity, and six-byte spin	Rails, circles, bumper, record 2, record 57, and record 72 have exact expected transitions in Rust and C tests	Verified
Timer ordering	1000:eab1 fixes 5/4/3/2/1 substeps, ball-slot order, countdowns, mechanisms, flippers, and Tilt decay	Timer-batch, collision-break, per-slot, target-rotation, panel, claw, countdown, and Tilt tests cover the recovered callback order	Verified
Attract mode	Raw 1000:01b7-063b distinguishes quotient and remainder outputs; marquee GDI lifetime is reconstructed	%32/2, %32/4, %40/4, %16/4, %144/8, magnetic %32, and six-callback marquee cycles are tested and rendered	Verified
Inputs	1000:638e/6bf8 define low-byte scans, Key-Up actions, auxiliary-window gate, launcher impulses, nudges, Tilt, and flipper latches	Both Ctrl/Enter aliases, both plus positions, release-time add/nudge/F12, held launcher, flipper rearm, and auxiliary-window blocking follow those paths	Verified
Rules and scoring	1000:b476, 1000:c4e1, 1000:bc36, 1000:ae6e, and 1000:9e85 define awards and state transitions	Bank progression, ninth-diamond order, ball-scoped double, marker comparison, lock transfer, special hole, all seven effects, Type-3 Tilt finish, and score/sound ordering have focused tests	Verified
Multiball	Records 148, 174, and 175 plus timer collapse/spawn paths define ownership and lifetime	Deferred spawn, per-slot capture age, survivor promotion, no-third-ball guard, nudge, fifth-lock deferral/reset, special respawn, and contact clearing are tested	Verified
Player turns	1000:ae6e saves all record activity/contact mirrors and advances live players	Four-player bounds, add-before-launch, per-player rule state, final-ball high-score ordering, flipper key lifetime, and normal/special drain behavior are tested	Verified
Audio	Sixteen WAV resources and every in-image play/stop call are inventoried	Samples are unchanged; monophonic async replacement, full volume, Tilt suppression, threshold precedence, panel stops, and no-WAV-2022 behavior are represented	Verified
High scores/configuration	The 276-byte IWIK table, dialog 32514, HighScore child, and INI reads are reconstructed	Signed-high comparisons, all ten entries, 21-byte name semantics, OK/Cancel flow, localized help, original settings names/defaults, and screenshots are tested	Verified
Randomness	The Borland LCG, high-product Random(n), Real48 result, and program-global seed are reconstructed	All gameplay draws use one exact stream; seeded traces are repeatable and the final seed continues across games	Verified
Stability/build	Native replacement must preserve finite state and compile on supported hosts	Production Linux build, 125 Rust tests, strict Clippy, two-minute finite autoplay, and Windows/macOS cross-target checks pass; native CI covers all three hosts	Verified

Differential evidence

• Record 2 special response: (157000,453000)/(0,3000) publishes the special ball at (17000,23000)/(0,3040) and continues to (16698,21191)/(-302,-1809).
• Record 57 left relative slingshot: (75530,354765)/(3000,15) becomes (77369,356597)/(1839,1832), Real48 spin -5.28 (83 5d 8f c2 f5 a8).
• Record 72 right relative slingshot: (235348,356825)/(-3000,15) becomes (233253,358494)/(-2095,1669), Real48 spin +5.46 (83 86 eb 51 b8 2e).
• Ordinary circle 155, the outer/inner shooter rails, bumper response, all four claw terminals, moving flippers, Type-3/4 records, scoring, and multiball also have focused C/Rust or Wine-backed tests documented in RECONSTRUCTION.md.

Deliberate host substitutions

• The native fixed window replaces ObjectWindows/Win16 ownership, but retains the title, logical client size, original pixels, controls, modal gameplay blocking, and visible flows.
• Portable JSON replaces writes to HISCORES.DAT and TDKPIN.INI; the original files are imported, while table ordering, names, settings, and UI semantics remain unchanged. F9/F10 expose additional viewers/settings without replacing any original gameplay key.
• Windows and macOS were cross-checked in this Linux checkout; native execution on those hosts is delegated to the repository's three-OS CI matrix. The semantic core is platform-independent and has no target-specific branches.

Modern additions and their divergences

Crazy mode (main-row 0) adds up to 64 live balls. To make every ball behave identically, the two original timer slots became one ordered ball list. One- and two-ball play keeps the recovered slot order, slot-numbered contact ownership, deferred effect-seven spawn, survivor identity, nudge draws, and multiball scoring, and all multiball tests above still pass. These deliberate differences follow from treating every ball alike:

• A ball resting on the shooter stop returns to the launcher from any slot (the original deleted a slot-two ball there) unless another ball already waits.
• Flipper kicks, record countdowns, and target rotation continue for balls in play while another ball waits in the launcher.
• Ball-to-ball contact ignores only the claw-held ball, and every ball's collision is published as the last collision record.
• When a ball leaves, claw and lock-hole contact owners move with their ball, and a lock hole the leaving ball was settling into is freed.
• The rim of a filled lock hole only repels balls from the outside, so a ball that was settling into it when another ball filled it rolls back out.
• The release-ball effect (effect seven) frees a parked special-hole ball whenever one is parked, not only when exactly one ball is in play. The released ball owns the hole's contact under its own slot number.
• The claw re-targets a shooter-lane release (terminal frame 18) to a table release whenever more than one ball is in play during its ride, not only when the capture began in multiball.
• High-score entries carry a crazy flag that both local storage and the shared service persist; older entries default to normal scores.

Reproducible final gates

cargo build --profile production
cargo test --workspace --all-targets --all-features
cargo clippy --workspace --all-targets --all-features -- -D warnings
cargo check --target x86_64-pc-windows-gnu
cargo check --target x86_64-apple-darwin
bash original/tools/test_reconstructed_c.sh
python3 original/tools/audit_reconstruction.py --require-complete

The C ledger gate reports zero raw, partial, restored, unknown, blocked, or unclassified required units. The only files intentionally outside commits are user diagnostic media in the repository root.