Drop the dead default-initializer on a definitely-assigned local

[richlander]

Found by the --compile-back oracle (#604).

A local whose defining store is not its first entry-block reference declares up front, and the emitter always spelled that int V_0 = default;. When the local is in fact assigned on every path before each read — the common shape of a value computed across switch sections or try/catch arms — that initializer is a dead store the original IL never had (locals lean on .locals init). Recompiling emits an extra ldc.i4.0; stloc and diverges from the source opcode stream:

orig : ldarg switch br ldc.i4 stloc ... ldloc ret
recmp: ldc.i4 stloc ldarg switch br ldc.i4 stloc ... ldloc ret   (leading dead store)

This declares such a local bare. A new conservative structured definite-assignment walk (ComputeReadBeforeAssign) decides: it threads an assigned set through if/switch/try-catch/try-finally/loops, joining on the paths that reach each merge, and keeps = default for any local read while not provably assigned. It under-claims assignment wherever unsure and bails outright — keeping every = default — on control flow it does not fully model (gotos/labels, leave, lock). So a bare declaration is emitted only when CS0165 cannot arise; the change can only remove a redundant initializer, never a required one.

Soundness: faithful. Dropping the initializer leaves the local to .locals init zero-initialization, exactly as the original relied on; the value is unchanged and the dead store disappears from the recompiled IL.

Measurements:

• Compile-back fixture corpus: exact 133 → 139 (+6), docket 20 → 14 (−6), recompile-fail unchanged at 498 (no CS0165 introduced — the soundness check).
• Compile-check arbiter: unchanged (semantic 316, Full malformed 1206 — no method-level regression).
• Tests +2 (428): PowerOfTwo/CatchEverything/TryFinallyAdd now declare bare; ParseOrZero (reached via a by-ref out-argument) still keeps = default.