refactor(daemon): give generic leaves their own execution vocabulary

[thymikee]

Summary

Part of #3276 (collocation). The largest import cycle in the repo was 33 files across src/daemon/interaction/internal and the daemon runtime leaves. One misplaced type held it together.

ResolvedGenericExecution and its parts were declared in request-generic-dispatch.ts, but only the leaves use them: runtime-admission, focus-runtime, scroll-runtime, and nine others. The dispatcher itself never used ResolvedGenericExecution. Each leaf type-imported its caller. Through the dispatcher's lazy import() of generic-settle.ts, that closed a loop through the whole interaction runtime. That lazy seam exists to keep the interaction runtime out of this cycle.

The four types now live in src/daemon/generic-leaf-execution.ts, below both sides. This is a type-only move: no runtime or bundle change.

Cycles, all edge kinds	main	this PR
Largest	33	7
Count	23	22

R9 only counts value and type edges, where the largest cycle was 6 files before and after, so no gate saw this knot. I'll propose extending it separately.

Validation

At 0920aa908:

• pnpm check:affected --run passes, including layering, fallow, and typecheck.
• pnpm depgraph before and after gives the table above.

🤖 Generated with Claude Code

[cubic-dev-ai]

No issues found across 24 files

_{View guided diff | Turn on auto-fix | Re-trigger cubic}

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	5.13 MB	5.13 MB	0 B
Package (unpacked)	5.13 MB	5.13 MB	0 B
Package (download)	1.54 MB	1.54 MB	+7 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	27.0 ms	28.1 ms	+1.0 ms
CLI --help	83.8 ms	86.7 ms	+2.9 ms

[thymikee]

This PR is ready at 2cc75f3. I found no problems in the diff, and there are no conflicts. The only check still running is Smoke Tests. The imports are all import type, so they are erased at runtime and the emitted JS is unchanged, which means the change cannot affect a smoke path. The size report also shows 0 B installed-size change. I did not re-run pnpm depgraph or pnpm check:affected, so the cycle counts (33 -> 7, 23 -> 22) and the layering, fallow and typecheck results come from the PR body. I also did not check the remark that R9 only counts value and type edges against the gate. Nothing else stands in the way once Smoke Tests finishes. Not blocking, and you can take or leave these: the doc comment at src/daemon/keyboard-runtime.ts:37 still points at ResolvedGenericExecution in request-generic-dispatch.ts, but that type now lives in generic-leaf-execution.ts. The header comment at the top of src/daemon/generic-leaf-execution.ts tells the import-cycle story, and AGENTS.md asks to keep that kind of narration out of comments, so it could shrink to one line or go.

[thymikee]

Took both points in 0920aa9. The keyboard-runtime.ts doc comment now names generic-leaf-execution.ts, and that module's header is one line stating the constraint. pnpm check:affected --run passes.

Evidence for what you didn't reproduce, from pnpm depgraph plus getStronglyConnectedComponents on the same graph.json:

	main 81f466640	this PR
Value + type edges: largest cycle / count	6 / 17	6 / 17
Value + type + dynamic edges: largest cycle / count	33 / 23	7 / 22

The first row is what R9 measures. On this branch the gate prints "the largest type-level cycle is 6 files (R9)", the same as main, so R9 never saw the 33-file knot. It only closes through the dispatcher's dynamic import() of generic-settle.ts.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-10-08 10:55 UTC