Claude Code claims a fix is live without verifying against the actual rendered output

Working through Claude Code (Sonnet 5) on a Home Assistant dashboard styling task. Pattern observed across ~1.5 hours / multiple restart cycles:

1. Claude edited a Lovelace dashboard card's HTML/Jinja content, asserted it would render correctly, and asked me to restart Home Assistant to check — without ever confirming the actual rendering behavior of the target frontend component first.
2. This repeated 3 times: (a) inline style="" attributes silently didn't render in HA's markdown card, (b) a Jinja {% set %} scoping bug inside a loop produced an empty table, (c) an assumption about Jinja whitespace-trimming behavior in the live environment (not verified) broke table parsing.
3. Each cycle cost a real Home Assistant Core restart (~1-2 min downtime) plus my own time reviewing screenshots and reporting back, because Claude's own local testing didn't match the live environment closely enough to catch the failure before asking me to check.
4. When asked about the cost of this, Claude could only "acknowledge" it verbally — it has no mechanism to log the issue, credit usage, or otherwise register the cost against the product, and directed me to file feedback manually instead.

Ask: local verification (or an explicit "I can't verify this without you checking" caveat) before asking a user to take a real-world action like a service restart. Separately, some in-product mechanism for registering repeated-failure cost/feedback would be a lot more useful than "log a complaint" resolving to "type it into GitHub yourself."