fix: stop discarding errors that had somewhere to go

[meszmate]

Stacked on #132 — the layers example needs a fallible view to try the compositor.

Item 6 from the review: an audit of the 194 catch {} sites in src.

163 of them are in component view functions, where the signature genuinely cannot report anything. Those need the fallible-view treatment (#132) applied component by component and are left for a follow-up. This PR is the remaining 20 — the ones where an error was being dropped despite having a caller that could handle it.

LayerStack.render returned a truncated frame

const grid = allocator.alloc(Cell, w * h) catch return "";
...
writer.writeAll(cell.content) catch {};   // × 5

An out-of-memory frame reached the screen as an empty or half-drawn one, indistinguishable from a compositing bug. render is fallible now.

Image commands dropped real I/O failures

.cache_image and .delete_image swallowed everything, while .image_file immediately below them propagated. An unsupported protocol already returns false rather than erroring, so what was being discarded was genuine I/O failure — and a cache_image that quietly failed left the following place_cached_image drawing nothing, with no signal anywhere.

An append that cannot fail

dev_console did clearRetainingCapacity() then appendSlice(keep.items) catch {}. keep is a subset of what was just cleared, so the capacity is already there. appendSliceAssumeCapacity states the invariant instead of hiding behind a catch.

The rest are best-effort, and now say so

Terminal cleanup and the suspend path run while already unwinding, with nobody left to report to — a broken pipe must not stop the remaining modes from being reset. Those keep their catch {} and gain a comment explaining why, so the next reader does not have to work it out again.

Tests

A failing-allocator test walks the failure index across LayerStack.render and asserts the result is always either a whole frame or an error — never something in between.

zig build test and zig build clean on 0.16.0.