docs(reverse): decode claw release arithmetic

Correct the coordinate model from guessed 16.16 values to the millipixel arithmetic shown by the 16-bit division routine.

Decode the Borland Real48 direction constants and record the exact position and velocity tables used by every claw terminal.

Test Plan:
- git diff --cached --check
- Compare constants against segment 1 disassembly at 9bca and eab1

original/MECHANICS_PROGRESS.md

@@ -42,17 +42,22 @@ renamed.
   destination is `(238,47)`; closing rows begin at source y=0 and opening rows
   at source y=148.
 - Claw capture/contact uses sound 2015; release uses sound 2016.
+- The release-coordinate projection is `(raw + 500) / 1000` for rendering;
+  decoded physics centers are `(258,78)`, `(270,94)`, `(275,97)`, and
+  `(325,92)` for the four selectable terminal frames.
+- The speed scalar is 3800 millipixels per tick. Its Borland Real48 direction
+  constants are decoded for all release cases in the readable C companion.
 
 ## Explicit pending evidence
 
-- Decode the runtime/FPU conversion used to project the claw's raw 16.16
-  release coordinates and non-frame-18 velocity into playfield pixels.
-- Recover or differentially measure the timer cadence used for each claw frame.
+- Differentially verify the recovered 10/20/30/40/50 ms timer choices against
+  live original captures; the Rust port intentionally uses the default 30 ms
+  claw cadence independently of render rate.
 - Transcribe the full fixed-point flipper impulse calculation, then compare it
   with the Rust floating-point collision response.
 - Validate the reconstructed sequences against deterministic Rust traces and
   captured frames from the original executable.
 
-Those four items are intentionally not counted as semantic parity. The readable
+Those items are intentionally not counted as semantic parity. The readable
 C file preserves their raw constants and labels the unresolved conversion so a
 later implementation cannot silently turn an estimate into claimed evidence.