Cut a roster line between graphemes, not between code points

[kevin9327]

What is wrong

oneLine in server/src/channels/text.ts says in its own docstring that it cuts "on code points so
an emoji is never split". That holds only for an emoji that is one code point. Most modern emoji
are not:

input	oneLine(input, cap) on main	what it renders as
🇰🇷 report, cap 2	"\u{1F1F0}…"	a lone regional indicator — a boxed K
👨‍👩‍👧 hello, cap 3	"\u{1F468}\u{200D}…"	the man, then a dangling zero-width joiner
👍🏽 ok, cap 2	"\u{1F44D}…"	the thumb with its skin tone dropped
1️⃣ first, cap 2	"1…"	a bare digit, no keycap

Every one of those was measured against main.

That line is not incidental. It is:

• what the sidebar draws under a channel (previewOf, MAX_ACTIVITY_CODE_POINTS),
• the generated channel title (summariseOne, MAX_SUMMARY_CODE_POINTS),
• and the excerpt the titler is shown (openingOf, MAX_EXCERPT_CODE_POINTS).

So a message that opens with a flag renders as a letter in a box in all three places.

The change

Segment with Intl.Segmenter(undefined, { granularity: "grapheme" }) and cut between clusters, so
what a person sees as one character is kept or dropped whole.

A string can never hold more grapheme clusters than UTF-16 units, so a line short enough to be under
the cap by .length is under it by clusters too and returns without segmenting at all. That is
nearly every line a roster draws, which keeps the expensive part off the common path.

The three caller constants were named ..._CODE_POINTS. The unit changed, so the names did — that
is the whole of the diff outside text.ts.

Verification

bun test server/tests/channel-one-line.test.ts

• With server/src/channels/text.ts reverted to main and only the new test applied: 5 fail, 4 pass.
• With the change: 9 pass, 0 fail.

The four that pass before and after are the guard against over-correcting: short text returned
untouched, control characters flattened and whitespace collapsed as before, plain text cut in exactly
the same place with the same ellipsis, and a single-code-point emoji kept whole. If this change had
moved the cut for ordinary text, those four would say so.

The emoji cases do not assert one hand-picked boundary. They sweep every cap from 1 past the end of
the line and assert that what survives is a whole number of the input's own clusters taken from the
front — a property no half-emoji can satisfy.

bun run --filter server typecheck and bunx biome check on all four touched files are clean.

bun test server/tests/channel-routes.test.ts server/tests/channel-titler.test.ts server/tests/channel-activity-input.test.ts gives 58 pass / 29 fail both on main and on this branch — identical. The 29 are Postgres connection failures on a machine with no database, not a regression.

Note

The CHANGELOG.md entry lands at the top of ## Unreleased, the same anchor as #450 and #452, so
whichever merges last needs a one-line rebase. Happy to do it on request.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[davidmckayv]

Deep-reviewed against live code (correctness, governance, no vendor/secret/scale issues). Composed build+tests green. CHANGELOG/format rebase on CI-validated substance.