[GenericJam]

Problem

Every Mob.Component handle_event/3 / handle_info/2 sent {:component_changed, id, module}, whatever the callback returned, and Mob.Screen.Server answered with paint(state, :none). That paint never skips: it renders, expands every component, and commits the whole tree through set_root.

FMutate's instrument is a native view that sends an event per key press, key release and knob move, and its handler returns the socket unchanged for key events. On a Moto G Power 2021, 6 presses produced 12 commits (RenderStats). Each cost 5-15 ms on the BEAM side and 10-20 ms of Compose recomposition on the UI thread, and that was the main source of jank left.

Change

• Mob.ComponentServer notifies the screen only when assigns changed (!==) and render/1 of the new assigns differs (!==) from the props the screen last fetched with render_props/1. A callback that returns the socket untouched, assigns the same value again, or changes only an assign render/1 does not read sends nothing. :rendered is read with Map.get and written with Map.put, so a component started before this change keeps working after mix mob.push.
• Mob.Screen.Server handles {:component_changed, ...} with repaint_if_changed/1, the fingerprint skip that forwarded messages already use. Several notices queued behind one paint now commit once. Navigation, activation, hot reload and :sync still paint unconditionally.
• The screen's own handle_info path was already checked. Since 2026-09-03 it skips the commit when the tree is unchanged, but it still runs render/1 and the expansion passes. That part is unchanged and recorded as out of scope in the decision.

Behaviour change

A component's render/1 is now treated as a function of its assigns, as in LiveView. Anything else it reads is picked up on the next screen paint, not by a callback that changed no assign. Identical props are also no longer re-sent to a native view. Screens have behaved this way since the 2026-09-03 decision. The moduledoc and CHANGELOG cover both.

Decision: decisions/2026-11-12-component-events-repaint-only-on-changed-props.md. I added an in-place note to 2026-09-03-skip-the-repaint-when-the-tree-is-unchanged.md, because its "only forward/2 may skip" no longer holds.

Tests

• test/mob/component_server_test.exs, describe "telling the screen to repaint", covers:
  • an unchanged socket, a re-assigned same value, and an unrendered-assign change: no notice
  • a rendered change: notice
  • the comparison baseline is the last render_props/1 fetch
  • handle_info and dispatch/3 follow the same rule
• test/mob/component_repaint_test.exs (new) counts set_root at a stub NIF, end to end through a screen with a native view:
  • an ignored event, an unrendered-assign change, and 25 ignored events: 0 commits
  • a rendered change: 1 commit
  • a bare {:component_changed} whose tree is already on screen: 0 commits
  • 5 changes queued behind a suspended screen: 1 commit

Mutation-checked:

• Reverting component_server.ex fails 6 of the new component-server tests.
• Reverting screen/server.ex fails 2 of the component-repaint tests.

Full suite: mix test, 2170 passed (28 doctests, 2142 tests), 38 excluded, 0 failures. mix format --check-formatted, mix credo --strict and mix compile --warnings-as-errors are clean, and the pre-push hook passed. A pre-commit adversarial review found nothing blocking.

Verification rung: host BEAM tests only. Not yet measured on a device.

[GenericJam]

Follow-up c68db05 (review fixes):

• Blocking fix: Mob.CrashReport.format_status/1 now redacts a component's :rendered, the props render/1 last returned. Before this, a component crash report or :sys.get_status/1 printed them in plain text, alongside assigns that were already redacted (MOB-310). Regression test in crash_redaction_test.exs; it fails without the fix.
• ecto 3.14.0 → 3.14.3 (security advisory).
• Hot-push regression test: a component whose state lacks :rendered still notifies and renders.
• Corrected the stale repaint comment. The CHANGELOG now says native views no longer get an identical-props re-commit after rejected input, and how to force one.

Device verification (fixture native component; root_renders counts screen renders):

	Moto G Power 2021 (Android 11)	iPhone SE 3rd gen
no-op taps ×5	no repaint (1)	no repaint (1)
hidden-assign tap	no repaint (assign changed, render didn't)	no repaint
rejected controlled input	no repaint, field keeps typed text	same
rendered change	1 → 2, lit=1 shown	1 → 2, lit=1 shown

Reviewer: reviewer subagent (Codex rate-limited until 9:31 PM): SHIP.
Host suite 2172 passing. CI: https://github.com/GenericJam/mob/actions/runs/43332023743 ✅