Here's an example, I guess:
Note that it uses MMBTH10s, which are "RF", though not too too different from '3904s. I mainly chose them for the low Ccb (low s12 if you prefer). The higher fT is of course a boon. The oscillator is 100-130MHz; if it were scaled down to 50 or 30MHz, the circuit would work very much the same with 2N3904s (in terms of equal h_fe).
Maybe not the best example, as a common-base amp is used for improved isolation; and the final (~CE) amp has low gain and wide bandwidth.
For example, the filter was tuned in part by varying R13 (which varies gain as well).
Performance was confirmed between measurement and simulation. The Thevenin output impedance of Q1 is very close to 10k || 1pF (represented by R1 and C1, as the part model didn't quite fit this parameter), and the filter values and Q2 input impedance were close as well (I forget what, something like 4pF || 400 ohms -- alas, I wrote it on a sheet somewhere).
I mean, if all you know how to do is make pretty circles on a Smith chart...
I guess you'll have to make pretty circles, and never do circuit analysis?
That's all you
can do, that's what you're asking for!
That, or do it empirically with a jig and two bias tees.
I'm sure many would be grateful for the confirmation, of having s-params for common devices. Especially for devices that are sometimes useful beyond their intended bandwidth. Like the aforementioned audio power transistors, or say, PN2369 (a switching transistor) used as RF preamp, or 2N4401 as medium power RF amp. Or noise figure data, at frequencies not normally covered, like 2N5486 at HF or LF (only VHF+ is documented), or an audio or switching JFET at HF.
So, you're not at all wasting your time, I think. But it will be a lot of time to nail things down tightly, and run several parts and types, and process and package the data.
Personally, I begin a design like the above pictured, by assuming the collector output is a Norton AC source: current, in parallel with a fairly high resistance and a capacitance. If grounded-base (or gate) or cascode, I ignore (AC) feedback (Ccb), and h_re is always quite small so that it isn't important at DC either. That leaves Zin, of the first base (cascode) or emitter (GB), which is modest or very low, respectively. (It doesn't take much base capacitance to turn the emitter impedance negative; it's unclear to me how you're supposed to match to this, so I ballast it up with a series emitter resistor, which also serves as termination.)
Collector load impedance is matched either to the small-signal collector impedance (maximum power gain), or the operating point (maximum power output). I think I designed the above filter as a one-end-terminated prototype, though it evolved into a strange mixture of things (I have the real PCB in front of me; it has a number of tweaks missing from the schematic), due to consideration of parasitics, and ultimately, by improved experience with filter design, in part because of this example.
This approach also works just fine for tetrodes and pentodes, if you're into that sort of perversion. The feedback term is quite small (usually fractional pF, even in sweep tubes), so it is only important at quite high frequency, or high gain (low bandwidth). (That said, I've seen 6AK5s needing shielding and neutralization -- they're not really IF tubes, and that circuit showed why.
) The grid input is capacitive and resistive, due to the combination of cold interelectrode capacitance (datasheet value), electron beam mass (about doubles the grid capacitance between operating/cold, for high-Gm and frame grid types), and transit angle (grid resistance -- example, E180F is around 1kohm equivalent at 100MHz and 10mA Ip).
Many datasheet values are inexact, or absent entirely (how often do you see excess grid capacitance, or grid conductance, specified? -- not often, and E180F is one of the few cases!), and real parts typically have enough variance that you can't possibly design a circuit, from scratch, with the most gain and selectivity possible, and no trimmers. That's why my prototypes have been refined iteratively, using the coupling/filter networks themselves as a probe to determine the amplifier's impedances, passing this data into SPICE, making further changes there, then implementing and confirming those changes IRL. This is a perfectly fine design method too, as long as allowance is made for component tolerances, and trimmers are added, when needed, and with enough adjustable range to accommodate everything.
Speaking of probe networks -- for this project,
https://www.seventransistorlabs.com/Images/FMRadio2.jpg I built pluggable tuning/filter/coupling networks,
https://www.seventransistorlabs.com/Images/FMRadio4.jpg including (not pictured) some wideband, low impedance dummy networks: a 50-ohm pass-through for the input, a balun for the mixer, and a balun (180° combiner) and a common mode (0° combiner) network for the mixer output (to measure balance, LO power, feedthrough and gain). Using all of them together, the system gain is a pitiful -14dB, though when you consider that's at 50 ohms (into a ~1k grid, then out of a ~10ka-a plate), that's pretty good for these poor tubes. The bandwidth is also something like 250MHz...
Incidentally, that 6J6 delivers harmonics all the way up to 800MHz. The LO is driving it nice and hard, and it's chopping pretty good. Hardly anything left up there, mind -- it's many dB down. But it's not nonexistent. It's defined by the physical limits of a tube of those dimensions, I think: where transit angle is falling over, therefore gain is nil, or even inverting. (Better is certainly possible -- see planar triodes -- which have far tighter interelectrode spacing, and proportionally higher limits. Even more still is possible, but only with distributed, free-electron methods -- TWTs.)
Tim