host: a text bound judges in UTF-16 units, so a cut hits text that is already under its own limit #2321

Implementation: FlyBot (ours) — PR follows within 24h. Do not pick up without a pickup-ok comment on this issue.

Written-by: FlyBot Lead
Claimed-by: FlyBot Lead
Claim-Id: ecea1e88081c47a7984f01947df78698
Branch: dev/4-fits-text-bound
Base: e40cecf
At: 2026-11-04T00:45:13Z

Implementation: FlyBot (ours) — PR follows within 24h. Do not pick up without a pickup-ok comment.

Written-by: FlyBot Lead

What is wrong

The host bounds text by UTF-16 units (text.length) and cuts it by code point (codePointEnd). A character that costs two units therefore eats two of the budget, so a bound can cut a text that is already under its own limit.

Measured on live main e40cecf6f5a5b0d19c95797792e98c98bf6382f5 (v0.2.22). The three files this issue
changes are byte-identical to 96eb3ef4458ca0d0659a9e4584b5bae792c3b8af, where the table was first
read: core/src/host/text.js blob ad6e627d, core/src/host/memory.js blob 5acc4a32,
core/src/shared/text-boundary.js blob dd411bc8:

A 300-character text is cut by a 400-character limit. The result reaches the user and the Bot itself through read_own_chat list / around / handoff lines (core/src/host/parts.js:344, :355, :465) and search rows (core/src/host/chat-index.js:213, :310), which makes the Bot fetch again what it already had.

clip (core/src/host/text.js:61) and excerpt (:47) are the two implementations with the defect; every row above is a consumer of one of them.

What to change

1. core/src/shared/text-boundary.js: export fits(text, max) — true when the text is at most max code points. codePointStart, codePointEnd and cut stay as they are; cut keeps flattening with toWellFormed().
2. core/src/host/text.js: clip (:61) and excerpt (:47) judge with fits, and keep cutting exactly as today (codePointEnd / cut). No other behaviour changes.
3. core/src/host/memory.js:1270 carries a third copy of clip with the same shape (text.length > max ? cut(text, max).text + ' …(cut)' : text). Replace it with the shared implementation — the module already imports codePoints from ./text.js at :30. Keep the printed suffix the same as today, or move to the shared … (cut) wording; that is the only text diff this issue allows.
4. New core/test/text-bounds.test.js: a table of "limit → consumer" rows that asserts each display/input consumer judges with fits. A future row that judges in UTF-16 units fails with its own file:line, so this list cannot silently rot.

Reproduction of the check itself

On the current tree the new test must fail first, naming core/src/host/text.js:61 (and :47) as unit-based, and core/src/host/memory.js:1270 as a duplicate. After the change the whole table passes and core/test/text.test.js stays green.

Acceptance

• node --test core/test/text-bounds.test.js red on the current tree with core/src/host/text.js:61 / :47 / core/src/host/memory.js:1270 in the message, green after the change.
• clip('😀'.repeat(300), 400, 'r') returns its input unchanged (300 code points ≤ 400);
  clip('😀'.repeat(201), 200, 'r') still ends with … (cut; r) and still keeps 200 UTF-16 units
  (100 code points) — the cut point stays where it is, because core/test/text.test.js:94 pins it
  ("keeps the first code-point-aligned prefix within 4000 UTF-16 units"). This issue moves the
  judgment to code points, not the cut.
• core/test/text.test.js and the rest of the core suite stay green; TZ=UTC npm run verify clean.

Scope

• In: fits(), clip, excerpt, the memory.js:1270 duplicate, the table-driven test.
• Out: every cut-point-only fix (a slice that can split a surrogate pair — its own issue), and the wire / admission / budget bounds in the audit comment. Those change what a peer sees on the wire or what a user is allowed to store, so they are judged separately.

The 33-row audit this scope comes from: PR #2000, comment 6608815594 — rows A (core/src/host/text.js:61, :47, parts.js:344, :355, :465, chat-index.js:213, :310).

Dedup

GitHub code search repo:ThinkFlowLab/FlyBot "export const fits" → 0 hits, so the helper does not exist yet. The closed family #1908 / #1912 / #1916 / #1917 / #1921 / #1957 / #2007 / #2022 are all single-point cut-point fixes (all closed, all in the cut half of this problem); none of them adds a judgment-unit helper or a check that stays red. No open issue matches (issue search is:open text bound clip excerpt code point limit host → 0).
The nearest closed family — #1907, #1912, #1918, #1925, #1984, #1993, #2006 — is all
cut-side (a slice that splits a surrogate pair); this is the judgment side (text.length > max).