prompts: add more details for the PCI subsystem

[kwilczynski]

The PCI subsystem has some specific needs and wants, so capture some of these for Sashiko. Perhaps with the extra content, Sashiko will be able to capture PCI-specific issues as part of the review process.

[rgushchin]

Thanks, Krzysztof! Overall I like the change and it might make reviews better. Few concerns though:

1. Who is supposed to do what? (Target Ambiguity)
   The prompt uses commands like "Add a changelog" or "Strip timestamps". Because the AI is reading this, it might get
   confused and think it needs to do those things, rather than checking if the patch author did them. We should change
   these to "Verify that the author..." so the AI knows its job is just to check.
2. Checking things the sashiko can't see (Unverifiable Rules)
   Some rules ask the AI to check if the author waited a few days before reposting, addressed all previous feedback, or
   updated Patchwork. Sashiko only sees the single patch file and has no access to email history as of now. If we tell
   the AI to check these, it will just guess (hallucinate) or complain about things it can't actually verify. We should
   remove these workflow rules for now.
3. Some redundancy in quick checks (maybe it's ok, but worth thinking of)
   The "Quick Checks" section at the end of the file is a copy-paste of almost all the rules we already explained in
   detail. This wastes tokens and means if we want to change a rule later, we have to update it in two places.
   We should clean it up to only include quick, unique checks.
4. Missing reviews for PCI drivers (Trigger Gap)
   The prompt only loads if the file is inside drivers/pci/ . However, network, USB, or graphics drivers also use PCI
   functions. If a network driver has a bug in its PCI setup, this prompt won't load, and Sashiko will miss it. We should
   make the prompt load whenever common PCI functions are used anywhere in the codebase.

[kwilczynski]

@rgushchin thank you for the feedback!

Some things that I didn't think about from the LLM angle. Let me incorporate some feedback from Bjorn, and then from you, and I will push revised document.

[kwilczynski]

@rgushchin let me know if this is any better.

[rgushchin]

It might actually hide some real issues. Not a hard blocker, just a note. It seems like there is some % of issues which AI can't reliably decide if it's real or not, so it just really depends if we prefer to miss some bugs or have less noise.

[rgushchin]

This all mechanics doesn't belong to per-subsystem prompts, let's keep it separately.

[rgushchin]

I think it's pretty much the same across all subsystems. Do we really need it here? Do you see any specific problems?

[rgushchin]

Thanks, Krzysztof! It's definitely better, but I still have some comments, so probably one more pass before I can merge it. Thanks for working on it!