Bodied Pascal cdecl procs: genuine SysV prologue (float params, >6 args)

Follow-up to bug-cdecl-indirect-over-6-integer-args. State after that fix:

• Indirect cdecl calls (dlsym'd C targets) marshal TRUE SysV including >6-int stack spill and xmm float classes — Synapse/OpenSSL 7+-arg entry points work.
• A BODIED pxx proc marked cdecl still receives the INTERNAL convention (every param in an integer register by position; >6 all-stack; floats as GPR bits). Direct pxx calls agree with that prologue, so pure-Pascal use works.
• The unsound overlap — p := @PascalProc into a cdecl proc-type when the proc has a by-value float param or >6 params — is now a LOUD compile error at the assignment (ir.inc AN_ASSIGN check). <=6 int/pointer params coincide with SysV and stay allowed (GTK-style callbacks).

Wanted: emit a real SysV prologue (and matching direct-call marshalling) for bodied cdecl procs: int/sse classes counted independently (an int param AFTER a float param shifts register!), xmm0..7 reception, >6/>8 stack reception at [rbp+16+...], Single narrowing at the right site (see the LANDMINE note in parser.inc's param homing about per-target narrow points). Then delete the AN_ASSIGN reject. Float RETURNS already work both ways (EmitLoadVar leaves xmm0 and rax both set).

Gate: cd8/cd11/cd12-style tests (8 int args, mixed float/int order, C-side callback via crtl calling a @PascalProc with double params) + make test + self-host byte-identical.

Triage 2026-08-19 (Track D re-triage pass, pin v363)

Genuine feature, still wanted — unchanged, verified by compiling.

• p := @MyCb where MyCb(a: Double; b: Integer): Integer; cdecl still stops at the loud AN_ASSIGN reject: "a Pascal-bodied proc with a by-value float parameter is not SysV-callable yet". The guard is intact, so the unsound overlap is still closed rather than silently wrong.
• A bodied 8-integer-argument cdecl function called directly from Pascal compiles and prints the right answer (36), i.e. the internal convention still agrees with itself. That is the ticket's stated state, unchanged.

Not a bug under the mandate's test: nothing produces a wrong value and the unsupported shape is refused at compile time with a reason.

Log

• 2026-08-30 — resolved, commit c5b8442e1.