build: line tables only and unpacked debuginfo, so the dev tree stops filling the disk

DWARF is ~97% of what rustc writes here: the largest `benilla-app` codegen unit measured 145 MB,
of which 61 MB was `.debug_str` and 13 MB `.debug_info` against 1.5 MB of `.text`. That carried a
1.5 GB `libbenilla_app` rlib, a 2.9 GB test binary and an 82 GB `target/debug`. `.debug_line`
survives, so a backtrace and `gdb bt` still name a file and a line; `split-debuginfo` leaves the
rest in the object files rather than copying it into the linked binary.

Cargo.toml

@@ -44,15 +44,28 @@ lua-src = { path = "third_party/lua-src" }
 kira = { path = "third_party/kira" }
 
 # Dependencies at opt-level 3 in dev: at opt-level 0, avian/parry collision drops a run to <1 fps.
+# Line tables only, and never packed into the binary, for the reason under `[profile.dev]` below.
 [profile.dev.package."*"]
 opt-level = 3
+debug = "line-tables-only"
+split-debuginfo = "unpacked"
 
 # Our crates at opt-level 1: Bevy's generic hot paths monomorphize into them, and at opt-level 0
 # Stormwind drops from a pinned 60 fps to ~45; the warm incremental rebuild cost is ~0.
 [profile.dev]
 opt-level = 1
-
-# No debuginfo trimming: `debug = "line-tables-only"` measures no recurring win on macOS + sccache.
+# Line tables only, not `debug = 2`. DWARF is ~97% of what rustc writes for this tree: the largest
+# benilla-app codegen unit measured 145 MB, of which 61 MB was `.debug_str` and 13 MB
+# `.debug_info` against 1.5 MB of `.text`. Carrying that cost a 1.5 GB `libbenilla_app` rlib, a
+# 2.9 GB test binary and an 82 GB `target/debug`, which filled the disk on a 1 TB volume and made
+# every write slower for it. `.debug_line` survives, so a panic backtrace and `gdb`'s `bt` still
+# name a file and a line; `split-debuginfo` leaves the rest in the object files rather than copying
+# it into the linked binary, which is what the link step paid for.
+#
+# This supersedes "no debuginfo trimming measures no recurring win", which was measured on macOS
+# with sccache and no disk pressure. Re-measure on macOS before assuming it holds there.
+debug = "line-tables-only"
+split-debuginfo = "unpacked"
 
 # The shipping build (`--profile ship`): fat LTO measures -1.0 cpu_ms (-5 %) against `release`,
 # which stays the measured profile because fat LTO's link time would tax every iteration.