[SizzleUnrlsd]

Fixes #166

Problem

The cross-TU summary of GlobalReadBeforeWrite keyed each global by its raw IR name, whatever its linkage. Two static globals with the same name in two files, or a static and an external one, shared one entry: a write to one counted as a write to the other. Reproduced on main faf1050:

Files	read_table (reads the static table of a.c)
a.c	0.55
a.c b.c (static table written in b.c)	0.50 + "at least one write ... in another analyzed function/TU"
a.c ext_table.c (external table written)	0.50 + same note

Fix

A summaryKey function now gives the key at the 3 places that read or write the index:

• an external global is keyed by linkerSymbolName, as the cross-TU function facts are (Cross-TU: rules that follow calls within a file lose them across files #157), so a declaration finds the definition of another file;
• a static global's key also carries its module. It stays in the index, so a write in another function of its own file still counts, but no other file finds it.

LLVM module identifiers are not unique here (every module is cached_ir or in_memory_ll), so the key uses the address of the llvm::Module. The index is only built in shared loading, from the modules that are then analyzed, so the address is stable for the lifetime of the index. The header documents the key.

Tests

• New check check_global_read_before_write_identity_across_files covers both collisions, a static global written in another function of its own file, and an external global written in another file.
• RED on the main binary: the 2 collisions fail, the 2 guards pass. GREEN: 4/4.
• Mutants caught: a static keyed by its name alone, and static globals left out of the index.
• CI-like suite on Linux x86_64: 2385/2385, unit tests included.

Before / after (main faf1050 vs this branch, Linux x86_64)

• Fixtures, default and z3: 846 runs, 0 differ, 1287 = 1287 diagnostics.
• Self-analysis, 4 chunks: 71 = 71 diagnostics, 0 differ.

Limits

• Fixtures are analyzed one file at a time, and the index only exists with at least two files. The self-analysis has no GlobalReadBeforeWrite result. The two comparisons show no regression elsewhere; the new check is what covers the fix.
• linkerSymbolName instead of the raw name only matters for a renamed symbol (__asm__). No case covers that here.

Code scanning

CI is green on 4c64dd4: 13 checks pass, the release attach step is skipped.

• 66 open alerts on the merge ref, the same numbers as on main: none new, none gone.
• No github-advanced-security comment on this PR. No new coretrace-concurrency-analyzer alert.
• This matches the local self-analysis before/after: 71 = 71 results, 0 differ.