logmessaged: drop records too big for the msgq queue instead of dying

[PeterPhuTran]

Replaces #312, which was auto-closed by an accidental force-push to its head branch (same fork-tooling mishap). Content unchanged.

Description

logmessaged re-publishes every log record onto the logMessage msgq socket. msgq
queues are 256KiB and msgq_msg_send() asserts that a message fits in a third of
the queue — so any record over ~85KB is a SIGABRT. The existing guard only rejected
records over 2MB, leaving an 85KB–2MB band where a single large log record kills the
daemon. And since a crashed logmessaged takes down the whole logging pipeline, the
crash itself can never be logged: the only symptom is a permanent
"Process Not Running: logmessaged" and silent loss of all swaglogs for the drive.

Hit in the field on 2026-07-10 (same device as #309/#311): updated logged a 130KB
git diff output record for a dirty working tree; logmessaged died seconds after
boot and stayed dead all day.

Fix: lower the guard to fit the actual queue capacity (MAX_PUBLISH_BYTES = 80KB).
Oversized records are still written to the on-disk log (the file write happens before
the publish) — they're only dropped from the msgq stream, exactly like the existing

2MB behavior.

Verification

New test_medium_log in system/tests/test_logmessaged.py, validated on a comma three:

• red on stock: a 200KB record kills logmessaged (msgq assertion)
• green with fix: logmessaged survives, later records still flow on the socket,
  and both records land in the on-disk log

Existing test_simple_log passes. Note: test_big_log fails on-device with and
without this change (it globs the live /data/log while a running logmessaged
rotates files — a pre-existing test-environment race; it passes in isolated CI).

Related: #311 flags logmessaged restart_if_crash as defense in depth; this PR
removes the crash trigger itself.

🤖 Generated with Claude Code

[github-actions]

Thanks for contributing! A maintainer will review your pull request soon.