fix: qdup/tuck/2swap bugs, getline, and cached word pointers

Co-authored-by: aider (openrouter/moonshotai/kimi-k2.6) <aider@aider.chat>

README.md

@@ -10,7 +10,7 @@ This is a toy/subset Forth interpreter designed for educational purposes. It pro
 
 - **Threaded Code**: The interpreter uses an inner/outer interpreter model. Colon definitions are sequences of word addresses (threaded code) traversed by an instruction pointer.
 - **Architecture**: 
-  - 32-bit signed integer cells (`int32_t`)
+  - 64-bit signed integer cells (`int64_t`)
   - Separate data and return stacks
   - Fixed-size dictionary, user memory, and compile buffer
   - Dictionary is a singly-linked list searched linearly
@@ -23,8 +23,8 @@ This is a toy/subset Forth interpreter designed for educational purposes. It pro
   - Memory access (`@`, `!`, `C@`, `C!`, `HERE`, `ALLOT`)
   - Simple string output (`."`)
 - **Limitations**:
-  - Fixed memory limits (dictionary, stacks, user memory)
-  - 32-bit signed integers only; no floating-point support
+  - Fixed user memory limit (1M cells); dictionary and stacks grow dynamically
+  - 64-bit signed integers only; no floating-point support
   - No file I/O or operating system interface beyond stdin/stdout
   - No immediate user-defined words or advanced introspection
   - Single-threaded execution