A lit lantern wasn't a refuge at all, just a place you died slightly slower. You were right that it should be somewhere you recover.

Now: stand still in a lit lantern for 1.6 seconds and it surges, filling you at ~22 a second. Twenty to full takes about four seconds.

from 20 warmth: 20 → 21 → 29 → 52 → 75 → 97 → 100

There's a ring that closes in while you wait, a warm swell of sound when it lands, and "WARMING UP" on the HUD. The delay is the point — it feels like the lantern waking up for you rather than a tap.

Every mode gained about nine points from it.

What the sweep said about embers

I tested your ember instinct properly, varying the count:

Supply isn't the problem — at 12 the bot spends two seconds of a fifty-second night in danger. What you don't have is information. The bot knows where all twelve embers are; you only see what's on screen. So instead of more embers, embers now glimmer much further through the dark, and the first time your warmth drops below 40 the game says the one useful thing: flare shows you every ember nearby. That was the actual answer and it was buried in a control list.

A bug I found while measuring

The same Hunted config on the same seeds gave 41%, then 66%. Three Math.random() calls were feeding the wisp wobble, which changes where wisps end up. My "reproducible" bot wasn't, and I'd been chasing five-point differences against twelve points of noise. Now seeded properly: two identical runs, identical numbers. Bonus — the shared-wood idea in DESIGN.md claims runs are reproducible from a seed, and that was quietly false. It's true now.

Where it landed

32 runs, same seeds, reproducible: