Cache common indentation in Inspect.Algebra

[dkuku]

Precalculate common indentation spaces (0..64) and collapsed newlines (0..2) at compile time.

Previously, indent(i) allocated two intermediate binaries per break (:binary.copy and @newline <> ...), creating immediate garbage on the process heap. Inlining @newline into the binary builder and returning compile-time literals for spaces(0..64) eliminates these allocations.

I chose 64 because :binary.copy/2 has $O(N)$ allocation and copying cost, whereas compile-time literals provide $O(1)$ lookups. Indentation can get quite deep—for example, with anonymous functions passed as the last argument—so covering up to 64 bytes yields the greatest savings while remaining within the on-heap binary limit.

Benchee for DateTime.utc_now() |> Map.from_struct()

Name                            ips        average  deviation         median         99th %
AFTER (cached spaces)      373.16 K        2.68 μs   ±361.58%        2.47 μs        5.46 μs
BEFORE (main)              334.00 K        2.99 μs   ±416.09%        2.72 μs        6.80 μs

Comparison:
AFTER (cached spaces)      373.16 K
BEFORE (main)              334.00 K - 1.12x slower (+0.31 μs)

Memory usage statistics:

Name                     Memory usage
AFTER (cached spaces)         7.87 KB
BEFORE (main)                 8.48 KB - 1.08x memory usage (+0.61 KB)

Assisted-by: Gemini 3.8 Flash

Minimizing allocation overhead in inspection paths is critical, as formatting large terms during error logging is a frequent cause of VM out-of-memory failures under heavy load.

[josevalim]

💚 💙 💜 💛 ❤️