[nicolas-abril]

comp.ts is 66,846 ttok, 846 over the 66,000 cap, so repo-gate fails until that is settled. Replaces #1356. It does the same job by inlining instead of minting specialized defs: +886 ttok against main, where #1356 cost +1,136, with the same speed.

What it does

When a call passes a lambda written in the source in a live parameter position, and the callee isn't recursive, the callee is inlined at the call site. The lambda is then applied in place, and no closure is built.

• Through a leading match: inlining passes through a callee's leading match when the matched type has one constructor (tuples, records). A small helper, def_at, puts the callback into the match arm.
• Beta: a lambda applied directly to its arguments is reduced in anf.
• Emitter: when a let binds the result of a one-arm match, the rest of the code moves into that arm. A match in tail position is emitted directly. This also applies to one-arm matches written in source.

There are no new defs, keys, markers or fuel. Termination comes from the callee being non-recursive.

Results

The fixtures from #1288 (by @mizchi) are included unchanged, and all 5 print their expected output.

Speed (M4, 10⁸ steps, 7 runs each, measured before the rebase onto #1968; after it, a re-measured helper-with-a-capturing-lambda loop went from about 0.26 s on main to about 0.01 s):

program	main	#1356	this
helper called each step with a fresh capturing lambda	241 ms	51.4 ms	51.4 ms
bind-style callback	841 ms	102.6 ms	102.7 ms
destructures a tuple literal	866 ms	26.3 ms	26.5 ms
destructures a call result (get-style)	911 ms	51.7 ms	51.5 ms

Closures built in the fixtures (main / #1356 / this, measured before the rebase):

fixture	main	#1356	this
known_callback	5	0	0
known_callback_array	2	0	0
known_callback_layouts	3	0	0
known_callback_shapes	11	3	3
known_callback_edges	21	18	19

Checks:

• test.ts passes 1803/1803 and perf.ts 124/124, on main at refactor(comp): shed tokens and compile large programs faster #1968.
• Everything under bench/ compiles to the same C as main, and JS output is unchanged across the corpus. The C changes in 42 of the 1,383 programs that load, including the demos app_slash_boss_3d and io_tcp_echos.
• tsc shows no new errors in comp.ts.
• repo.ts fails only on the ttok cap.

Not covered

• Matches with more than one arm (Bool, Nat, List helpers): the rest of the caller would have to be copied into every arm.
• Recursive callees.
• Callees with more than one non-callback argument left when the match is reached.

None of the fixtures need these.

Cost

+886 ttok. Getting under the cap needs about 846 tokens cut elsewhere in comp.ts, or a raise.

Not run: CUDA, Linux and the JS benches.

🤖 Generated with Claude Code