refactor(daemon): inline the touch sub-switch into the interaction dispatcher

[thymikee]

Summary

interaction-touch.ts was a 31-line pass-through: a second command-name switch that only
re-dispatched press/click/longpress/hover/fill to the same handlers
handleInteractionCommands already pulled in. Inlined those five cases into the existing
switch in src/daemon/interaction/internal/interaction.ts and deleted interaction-touch.ts.
No client-, operator-, or command-author-visible behavior change — this only removes one
indirection hop on the press/click/longpress/hover/fill dispatch path (−1 hop).

Review follow-ups landed in the same branch:

• The ADR-0011 responseData guard scanned files by an interaction-touch*.ts name prefix,
  which stopped covering the switch once it moved into interaction.ts. The dispatcher is now
  held to a structural rule instead: the guard parses interaction.ts with oxc-parser and
  requires every press/click/longpress/hover/fill case to be exactly one
  return await <handler>(...) whose callee is imported from an interaction-touch*.ts module.
  Nothing can be constructed in the switch at all, so a nested switch, a local responseData,
  an inline literal or a foreign callee all fail. The guard's own tests plant each of those
  shapes in an inline dispatcher source and assert the exact violation.
• docs/adr/0019-end-state-hop-trace.md traced both files this PR merges by name and pinned
  itself as unaffected by this PR. Merged the corresponding table rows, renumbered the
  remaining hops, and updated every downstream count/range reference in that doc and in
  docs/adr/0019-request-bound-platform-runtime.md.

Validation

Tested at de5036093f (structural guard), 0b982bfcc0 (oxfmt only) and 31af37752b (folds the delegation check under fallow's complexity threshold, then merges main for the #2310 size-report schema) on top of 292c0edb220e50f5f34f9436c777fafad631eba0.

• npx vitest run src/daemon/interaction/internal/__tests__/ — 30 files, 197 passed
• npx tsc -p tsconfig.json --noEmit — clean
• npx oxlint on the changed test file — clean
• Planted red (de5036093f): gave press a nested switch followed by a hand-rolled
  responseData in the real interaction.ts; the guard failed with
  case 'press': the case body is not exactly one return await (...)``, then passed clean after
  reverting. The same shape, an inline response literal, a non-touch callee and a removed
  command are pinned as permanent inline planted-red cases in the guard test (7 tests green).
• Full affected gate: green at 292c0ed (check:affected; format, lint, typecheck, layering, fallow, build, vitest-related)

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	4.44 MB	4.44 MB	-88 B
Package (unpacked)	4.44 MB	4.44 MB	-88 B
Package (download)	1.32 MB	1.32 MB	-13 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	26.7 ms	27.9 ms	+1.2 ms
CLI --help	80.7 ms	82.2 ms	+1.5 ms

[thymikee]

Sentinel review at 292c0ed.

P2: the new case-body regex stops at the first nested case/default, so a hand-rolled responseData after a nested switch in press is invisible to the construction guard; the previous whole-file scan caught it. The delegation move itself is behavior-preserving, but the filename exception plus copied command list weakens the invariant. Keep construction enforcement at the owning typed/delegation boundary or reuse structural parsing instead of formatting-sensitive case extraction. Plant a nested switch followed by hand-rolled responseData and prove the gate rejects it before restoring. Hold readiness until this is resolved.

[thymikee]

Addressed in de50360. The case-body regex is gone; the guard now parses interaction.ts with oxc-parser and requires each touch case to be exactly one return await <handler>(...) whose callee is imported from an interaction-touch*.ts module, so construction enforcement stays at the delegation boundary and nothing formatting-sensitive remains. Planted the nested-switch-then-hand-rolled-responseData shape in the real press case: the guard rejected it (case 'press': the case body is not exactly one return statement); reverted. That shape plus an inline literal, a foreign callee and a missing command are pinned as inline planted-red cases in the test itself.

[thymikee]

Sentinel review at 0b982bf: The nested-switch enforcement hole is fixed by the AST delegation check, independently reviewed. One actionable gate failure remains: delegationOf at interaction-response-construction-guard.test.ts:105 is 11 cyclomatic / 9 cognitive and fails Fallow in Compatibility & Provenance (job101366047914). Simplify/split the owning AST shape checks without raising the baseline, retain the planted negatives, and rerun the gate. Size also fails from incompatible base-report js.rawBytes; device CI must clear separately. No readiness label.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-09-05 20:41 UTC