feat(skill): add Dogfood (bug hunt) mode to the companion skill

[warrenbryant-msft]

What this adds

A third "Dogfood (bug hunt)" mode in the companion skill, alongside Smoke Test and Run my plan. The agent systematically explores an app to find bugs, UX issues, and polish problems, and produces a severity-ranked issue report (findings.md) with repro evidence for each finding.

This came out of a chat with @warrenbryant-msft about folding his "dogfood" QA approach into PowerAppsControl. The driving/recording is already your engine's strength; this layers on the QA judgment — what counts as a bug, how severe, how to prove it, and when to stop.

Why it's not just "Smoke Test" or the old "Explore"

• Different deliverable from Smoke Test. Smoke Test's artifact is a re-runnable test plan ("what the app does"). Dogfood's artifact is an issue report ("what's wrong, with proof"). The skill routes "find bugs / QA / what's broken" to Dogfood and "give me a test I can rerun" to Smoke Test.
• Not a revival of the removed standalone Explore mode (1.5.0). It's a QA/triage output, not free-form exploration.

What's in the mode

• Issue taxonomy — severity levels + categories, adapted to Power Platform surfaces (canvas non-response, model-driven grid/subgrid/lookup failures, delegation limits, permission banners, the "green click that didn't take effect" case you already flag honestly).
• Repro-first evidence, sized to the bug — interactive bugs get before/action/after screenshots + a timestamp/step marker into the existing session.mp4 (no per-issue video, since you already record the whole session); static bugs (typos, clipping) get a single annotated screenshot.
• Incrementally-written findings.md — created at the start and appended per issue, so an abandoned/interrupted session still yields a usable report.
• Stop rule (5–10 well-documented issues; depth > count) and an on-demand wrap-up.
• Dogfood -> regression bridge — offers to convert a confirmed bug's repro into a saved save_test_plan so "Run my plan" can re-check the fix.

Scope / decisions I left for you

1. Kept it as a mode, not a separate skill. It reuses Step 1 (open & verify) and Step 3 (session lifecycle) verbatim; only the middle is new. Easy to split into a standalone /PowerAppsDogfood skill if you'd rather — your call.
2. Deployment: inlined the taxonomy + report template as appendices in SKILL.md. I noticed InstallSkill copies only skill/SKILL.md (not the folder), so separate references//templates/ files wouldn't reach users. Inlining keeps this a single-file, skill-only change that ships through the existing --register pipeline with zero code/build changes. If you'd prefer the progressive-disclosure structure (separate files) I'm happy to follow up with the small InstallSkill (copy the skill/ folder recursively) + .csproj (glob skill/**) change — I just couldn't build-verify C# here (no net10 SDK), so I kept this PR code-free.

Files

• src/PowerAppsControl/skill/SKILL.md — adds Mode 3 + Appendix A (taxonomy) + Appendix B (report template).
• CHANGELOG.md — [Unreleased] entry.

No server code touched. Take it or leave it — no worries either way. 🙂