I'd check the solder joints closely. Another option is to measure the resistance to ground and VCC for the pins. If an input or output is bad it can be found that way. AFAIK there are no schematics for these machines but I managed to fix several problems in a similar HP RF generator using the block diagram from the service manual.
For the sequencer area, the "service manual" shows two rectangles with one line connecting them. One rectangle says "Sequencer" and the other says "Sequence RAM (64k)". Extremely lame. Plus, the sequence SRAM is actually 128k (Cypress CY7C109D), so even that isn't right.
I too thought it could be mechanical or a solder joint. I spent a fair amount of time poking at the board with no correlation.
The problem was a bad SRAM chip (U320) in the sequencer. This particular chip is responsible for one bit of the segment address, the two sequence markers, four bits of the repetition counter, and the end-of-sequence flag bit. In other words, a little of everything across the total of 32 bits output by the sequencer to really make a mess out of the waveform and spread the suspicion across many related components.
I found the problem by noticing there was only one write to the sequence SRAM to address 0 on waveform startup, yet the sequencer was looping through multiple addresses. The other addresses contained garbage because they had not been initialzed. Only address 0 had been set up properly by the firmware.
Further investigation showed that the end-of-sequence bit would properly be set to 1 at address 0 when the waveform was initialized, but then the bit would randomly revert back to a 0 within 30us to 500us. This caused the sequencer to play out the uninitialized memory beyond address 0.
It seems likely other bits and/or addresses had this issue also, which explained random changes in the waveforms I started to notice.
Occasionally, the bad SRAM would settle to a condition where it would repeat the proper segment, and the unit would appear to work. However, if I had loaded an I/Q pattern with any complexity, it would have certainly failed to play properly. The built-in multi-tone waveform I was using only needed a simple single-step sequence, so chances were much better it could fail, repeat this one sequence, and appear to be ok.
One take-away lesson: I bought this unit after checking there was a service manual for it, as I do for most used equipment from ebay. I did not discover it was only a Service *Guide* (a.k.a. Assembly Level Service Manual), until I really needed it.