Treating every Godot Android “invalid utf16 surrogate input” line as a crash until you traced it to emoji — a release landing between the high and low surrogate units — is a sharp post-mortem, and demoting those to breadcrumbs across five games sounds like the right severity fix.

Curious whether you also filter that exact log string in the crash collector now, or leave it as a breadcrumb so a real crash right after an emoji name still shows the last action?

Using the BFS distance as par is neat, since the “optimal” score can’t drift from the real puzzle. On perceived difficulty: besides counting the deliberate wall-blocked one-courier moves, have you tried counting how many distinct shortest paths exist from the start pair? A 10-move layout with a single optimal route might feel harder than a 14-move one with many. Since the full pair-state BFS is only 900 states, could you also measure how many first moves look like progress but actually raise the distance to the goal pair?

Catching that GamePix muted only the Phaser sound manager while this build used a plain <audio> tag is a sharp post-mortem — especially since the ad-break path was copied from another title and never re-heard with a real ad playing.

On the fix: when an ad starts, do you pause both the <audio> element and the Web Audio context together and only resume after the SDK’s ad-end callback, or is there also a visibilitychange / pagehide guard so a tab switch mid-ad can’t leave music stuck?

Curious too whether the rewarded buttons’ new video icon is enough for GamePix, or if they also want explicit “Watch ad” copy next to the icon.

Keeping the original 1996 generator and only changing the tile→3D builder output is a clean split — especially with the fixed 1 tile = 2 m scale so layouts stay familiar. Curious how the builder resolves T-junctions and corners when an arch and a wall share an edge: whole prefab swaps, or rule-stitched wall segments? Also, with the HUD still authored in 640×480 and one scale factor for ultrawide, did anything feel too sparse at 5120×1440, or did that single factor hold up across the demo?

Making ESC reliably pause sounds small but it’s the kind of fix players notice immediately, so nice to see it in 0.5.3. Was the pause issue related to the parallel map generation you described in the ledge-spawn devlog, or something separate in the menu handling? Also curious whether the Breached and Cracked stack cap mattered much for balance once they could actually reach max stacks.

Hiding among 100 NPC deer with up to nine player deer sounds like the crowd itself becomes the main mechanic, so I can see how a performance pass snowballed into reworking the deer too. Did the NPC deer behaviour have to change so player deer can genuinely blend in, or does the Hunter mostly spot players by how they react, like bolting when a TrailCam catches them? Also curious how the 40-second Wendigo night phase changed the pacing of a typical match.

Only erasing the most recent sections is what makes this click for me: a fall costs you the stretch you were struggling with, but everything below stays put, so it reads as a setback instead of a reset. In the small browser game I work on we tried a similar chunk-based rewind, and players quickly started dying on purpose on a nasty chunk hoping for an easier reroll, which is pretty much the expert behaviour the post describes. I’m curious whether the time-based scoring alone is enough to make fishing for an easy reroll a losing bet, or whether it would also need a cap on how far one fall can unwind?

Fun effect. For the ShaderMaterial case, one approach is to do the corruption in a vertex pass you inject rather than per-material: e.g. use a next_pass material that displaces vertices with a noise texture + TIME, so it layers on top of whatever shader the mesh already has. Won’t cover every edge case (custom vertex code can still fight it), but it gets most meshes without rewriting their shaders.

Classic overlap trap. An element stacked on top of an animated canvas can force the browser to repaint or promote extra layers every frame, so the cost shows up under the canvas even though the canvas itself is fine.

Your takeaway is the real one though: one change per analyzer run. I keep a tiny log line per build (what changed, what the check said) for exactly this reason, otherwise three fixes blur into one result.

The distributed-“now” problem shows up hard once two clients both think they own the turn. Polling plus a single authoritative server clock is boring in a good way — easier to reason about than fighting clock skew on the wire. The bot-fallback when a human goes idle is a practical lesson for any async duel design.

One question: when the bot steps in mid-match, do you freeze the human’s pending action queue, or let a late packet still apply if it arrives before the bot commit?