fix: deoptimize math-specialized functions on non-numeric arguments

[esrrhs]

Problem

A file-level math-specialized function whose math-position argument was non-numeric (string/bool/nil/table, including nil filled in for a missing argument) hard-failed directly in the TCC/GCC entry dispatcher. The exception escaped the generated dylib with no handler → terminate due to uncaught exception, aborting the whole process. It could not be caught by pcall, while the INTERP backend executed the function body normally:

function f(x)
    if type(x) ~= "number" then return -1 end  -- legal type guard
    return x + 1
end
f("abc")  -- interp: -1 ; TCC/GCC (master): process abort

Fix

• Every specialized function now gets an additional <fn>__dyn full-dynamic variant compiled with CompileFuncBody(..., spec_bitmask=-1) — the mature non-specialized CVar path. The entry dispatcher routes to it when any math argument is neither VAR_INT nor VAR_FLOAT, instead of throwing. Genuine arithmetic errors inside the body still take the pcall-able Lua error path, and all three backends now report the same body-level message (attempt to perform arithmetic on non-numeric value).
• New FakeluaSpecMissException (util/exception.h): the 7 tentative native-numeric compile sites (CompileNumericExp, raw native binop/unop) throw this instead of a logged hard error. TryCompileNativeExpr catches it as routine fallback; if it escapes Build(), Generate() converts it to the same Code generate failed, ... text as before.
• __dyn is forward-declared in the decls section.
• Recursion inside __dyn re-enters through the VM entry dispatcher; RTLD_LOCAL keeps same-named symbols isolated across files.

Tests

• New test/lua/infer/test_spec_deopt.lua + infer.test_spec_deopt: type guard, multiple math params, heterogeneous return (string), pcall, recursion, C++ boundary catch — across TCC/GCC/INTERP × debug on/off.
• exception.math_param_non_numeric_error updated to the unified message and now runs on AllJitTypes().
• InferGetCCode strips __dyn bodies by default so the 100+ existing shape assertions keep targeting the specialized variants; the deopt test opts in with keep_dyn_variants=true.
• Diff tests: 32 new cases vs Lua 5.4.9, 0 disagreements; prior 223 cases show no new regressions.
• Full suite: infer 220/220, exception 105/105, all pass except the 8 test_mysql.integration_* tests that require a live MySQL server (pre-existing environment baseline).

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 85.00000% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 88.00%. Comparing base (2c89572) to head (30f7ba1).

Files with missing lines	Patch %	Lines
src/compile/c_gen.cpp	84.61%	6 Missing ⚠️
❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #30      +/-   ##
==========================================
+ Coverage   87.99%   88.00%   +0.01%     
==========================================
  Files         124      124              
  Lines       25388    25414      +26     
==========================================
+ Hits        22339    22366      +27     
+ Misses       3049     3048       -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.