Three bugs this week had the same shape: they looked fixed locally and were still broken in production, because nobody had actually checked.
The chat advisor’s backend, /api/suggest and /api/usage, had three commits sitting unpushed. Every test I ran locally passed because the local server had the code; production didn’t, so real requests 404’d silently. Watch image search failed for a similar reason: Google closed its Custom Search JSON API to any project created after 20 January 2026, and Pulso is a July project, so the fix, swapping in SerpAPI, was necessary but I hadn’t actually confirmed the new key was set on Railway until I checked.
The pattern I’m taking from this: "it works on my machine" isn’t a finished task, it’s a hypothesis. A feature isn’t done until it’s verified against the real deployed environment, not the sandbox.
Smaller fixes from the same sweep: suggestion prices anchored on the wrong reference (one Tudor came back at £9,800 instead of a real £2,700-4,300), and a question about naming your own watch that silently burned a paid lookup for no answer.
