[frankmcsherry]

The join, half-join and reduce backends no longer compile their lambda a second time at the same input before narrowing it.
On the LDBC SF1 one-off replay this takes 70 µs off each IC read's install (757 to 687 µs), with identical answers.

Why

slot_graph compiles every slot at the shapes it reads.
The three backends then ran corgi::compile again on that graph, at the same input, before corgi::narrow.
A profile of one-off installs put the second compile plus narrowing at 140 of IS2's 462 µs install, and 177 of the average IC read's 757.

What changes

• slot_graph returns the corgi::Compiled it makes.
• compile_projection and compile_reduce return that Compiled, so the join, half-join and reduce backends narrow it directly.
• The map, filter, flatmap and join-function paths take its graph, as before.
• Tests that compared or built from the returned graphs now take .graph().

Compiling again at the same input changed nothing.
A temporary check, not committed, compiled every slot a second time during the interactive test suite and compared the graphs: all 53,801 were equal.

Measurements

One run, mini-02, one worker, diagnostics on, against a stored run of a6b7922a with the same uncommitted install timers on both sides.

	before	after
install per IC read (weighted)	756.7 µs	686.8 µs
render phase, IC1 / IC8	967.6 / 259.7 µs	797.9 / 212.9 µs
server CPU outside operators	14.3 ms/tick	13.8 ms/tick
whole run, instructions	2,130.0 G	2,121.7 G
whole run, server CPU	123.29 s	121.80 s

Answer and physical digests are identical on all 808 ticks.
Per IC family the saving ranges from 25 µs (IC13) to 171 µs (IC1).

Tests

cargo test -p interactive: 294 passed, 0 failed, 9 ignored; no warnings.

Implemented and measured by Claude Code.

🤖 Generated with Claude Code

https://claude.ai/code/session_01H6wB4H99TLw4tqu4x48pXL