Fix DCC GDB/MI debugging integration

[gloveboxes]

Summary

Improve ntvcm's DCC GDB/MI integration after end-to-end testing with the current MIR-based DCC emitter and VS Code's OpenDebugAD7 adapter.

This fixes source lookup and stepping edge cases, supports line-oriented CP/M input while debugging interactive programs, handles runtime-sized VLA watches, and substantially expands the cross-toolchain debugger regression suite.

User-visible fixes

• Emit a real absolute MI fullname for source records while preserving the recorded file value.
• Resolve relative source records against the loaded program location.
• Match source paths case-insensitively on Windows and macOS, including uppercase CP/M/editor names such as CHATC11.C against chatc11.c metadata.
• Avoid carrying a preceding function's closing-brace source record across a shared function boundary when stepping into the next function.
• Display runtime-sized VLA types as [] and allow indexed watch access when the static outer dimension is unknown (0 in DCC metadata).
• Require exact --interpreter=mi / -interpreter=mi option spelling so similarly prefixed CP/M arguments cannot accidentally start the MI server.

Interactive CP/M input

The Debug Console is owned by GDB/MI, so plain text is normally treated as a debugger expression. This PR adds an explicit target-input bridge:

-exec input 10*3
-exec input /bye

When a CP/M program requests input through blocking console input or a BDOS 6 polling loop, ntvcm reports a normal step stop. -exec input <text> then:

1. queues the text plus CP/M Return;
2. resumes execution atomically; and
3. runs until the next breakpoint, input request, or program exit.

This was validated against the interactive DCC CALC app, including breakpoints immediately after the BDOS polling loop.

Normal-mode isolation

• Target-input routing is gated by g_miMode.
• x80.cxx and x80.hxx are unchanged by this branch.
• The emulated instruction loop and cycle accounting are untouched.
• Normal-mode CPU diagnostics, output, and instruction cycle totals match clean main.

Regression coverage

Expand tests/debugger_mi.py to validate both emitted metadata and debugger behavior:

• strict parsing of every DCCDBG 2 record class;
• linked code address bounds, function ranges, variable lifetimes, struct IDs, flags, and dimensions;
• source breakpoints and duplicate-address behavior;
• absolute and relative program/source paths, dotted directories, quoting, and metadata capacity warnings;
• multi-module relocation order and aggregate globals;
• nested source lines and shared function-boundary stepping;
• recursive frames and frame selection;
• scalar locals, globals, pointers, function pointers, arrays, VLAs, structs, unions, bit-fields, lexical shadowing, and variable assignment;
• target input through direct MI and VS Code OpenDebugAD7;
• adapter disassembly and instruction stepping;
• exact MI option activation.

The old fixed entry-byte assertion was also replaced with a protocol-level four-byte hex assertion so the MI memory-read test does not depend on a particular DCC backend encoding.

Validation

Passed locally on macOS:

• ./mmac.sh
• python3 tests/debugger_mi.py --dcc-root /Users/dave/GitHub/dcc --ntvcm ./ntvcm
• the same suite with VS Code's installed OpenDebugAD7 via --adapter
• the direct MI suite under AddressSanitizer and UndefinedBehaviorSanitizer
• python3 tests/cyclecheck.py ./ntvcm
• CPUTEST.COM (CPU TESTS OK)
• sustained normal-mode 8080EXER.COM comparison against clean main: identical output and 23,835,739,337 emulated cycles

Notes

The test suite remains cross-toolchain by design: it builds compact DCC fixtures, loads their generated .COM/.DBG pairs into ntvcm, and exercises the same MI protocol used by VS Code.