Fixed: builds lost to problems we could have recovered from

We audited every build that failed in production over the last week. All of them were recoverable, and none of them recovered. This is the round of fixes that came out of it.

Builds that failed on something trivial

• An icon that doesn't exist can't break your site. The AI asked for an icon named Fire; the icon set only has Flame, and the whole build failed on it. The self-repair loop should have caught this and was blind to that particular class of error. It isn't anymore, and separately, unknown icon names are now mapped to the closest real one before the build even runs — offline, deterministic, no second guess needed.
• A GitHub outage no longer costs you a site. By the time we save your site's code, all the work of writing it is already done — so a single failed request there used to throw away a complete site. Every write now retries through transient failures. Two owners lost sites to this; one never came back.
• Setting up booking, a blog or a shop works before you publish. On sites that were still on their temporary address, those setups failed outright. They don't.
• A build that gets killed outright is recovered. It used to sit at "generating" behind a spinner that never resolved, because nothing was left running to notice.
• You are never billed for an attempt that failed on the way to the site you actually got.

Sites that told the truth about themselves

• No invented contact details. One published site carried two different phone numbers, neither of them real. Phone numbers and email addresses that you never gave us are now stripped before publishing.
• No menu links to nowhere. Navigation items pointing at sections that were never written are removed instead of scrolling to nothing.
• The mobile menu doesn't crash. A specific bad combination in the header could throw the moment someone opened the menu on a phone. It's repaired automatically, and the thing that caused it now steps aside.