Addressing DogFooding experiences feedback

[sdevinenamd]

Thank you @amd-hasnasir and Matt for testing out the skill and sharing your valuable feedback. Addressed the pointers in this PR.

Hasan's feedback

• SSN test prompt caused Claude to refuse → Since this is a model behavior and not consistently observed, added an alternative email-only backup command (Step 5). Fyi, the walkthroughs are only for internal devs so this would eventually become obsolete to the public.
• Steps 5/6 didn't say to use a new session → added "On a new Claude session, run:"
• Walkthrough expected rule id pii-stays-local, skill defaults to rule-1 → fixed walkthrough to match actual default

Matt's feedback

• "Write better prompts" doesn't fix silent misrouting - it's a judge-capability issue → router.model now defaults to most capable candidate, not cheapest. Also added a note to switch to smaller models if performance capacity is known
• Verification step was buried/undersold → promoted to mandatory Step 9
• Classifier label typos silently score 0 forever, no error → documented + added a verification check
• Validator can't confirm models exist on target host → noted in Step 1 / Step 8b
• Mode A trace ids are synthetic (__route_0, …), not author-given → documented in reference.md
• Empty rationale isn't a reliable fallback signal → clarified default_used is the real one
• GPU multi model resource gap → Lemonade allows loading >1 models depending on the GPU availability. However, clarifying that level of technicality is out of scope for this skill.

[danielholanda]

Changes look good