host: shorten a long message for the steward on character boundaries #2546

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: cfab780181c7410d941e05ea45bc8502
Branch: dev/5-steward-shorten-chars
Base: d400edb
At: 2026-11-05T14:36:23Z

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

Written-by: FlyBot Lead

What happens

The long-message view the steward reads still cuts on UTF-16 units. core/src/host/steward.js:48-51:

function shorten(text) {
  if (text.length <= STEWARD_LINE_MAX) return text
  const head = Math.ceil((STEWARD_LINE_MAX * 2) / 3)
  return `${text.slice(0, head)}\n[… ${text.length - STEWARD_LINE_MAX} characters left out of this view …]\n${text.slice(head - STEWARD_LINE_MAX)}`
}

Two failures, both from slice/length counting units:

• a cut at head (1000) or at head - STEWARD_LINE_MAX (the kept tail) can land between the halves of a surrogate pair, so the steward is handed a lone high or low surrogate — U+FFFD once encoded;
• a message of 1,500 characters whose non-BMP characters make it more than 1,500 units is shortened even though it is inside its own limit, and the marker under-reports what was left out.

#1993 / #2000 fixed five bounds in core/src/host/ on cut() — one of them steward.js:148, the checkpoint note's quoted first line — but shorten() was not in that table, so this is the remaining site of the same family. core/src/shared/text-boundary.js already carries the rule this should follow (fits, cut, codePointStart, codePointEnd, and the #2321 note that the budget is counted in code points).

Reproduction

Clean base d400edb74798b71e233eccd46f221eaef782ff5d, only the new case added to core/test/steward.test.js:

# Subtest: a long message shortened for the steward cuts on characters, never inside one
not ok 11 - a long message shortened for the steward cuts on characters, never inside one
  error: |-
    a character at the cut stays whole
    false !== true
# tests 11 / # pass 10 / # fail 1

Two shapes: 'a'.repeat(999) + '🚀' + 'b'.repeat(800) (head cut at 1000 lands inside the emoji → lone high surrogate) and 'a'.repeat(1000) + '🚀' + 'b'.repeat(499) (tail cut lands on the low half; this one is only 1,500 characters and is shortened anyway). Run: node --test core/test/steward.test.js.

Expected

shorten() uses the shared rule: at most STEWARD_LINE_MAX characters (fits), both cuts land on code point boundaries (codePointStart / codePointEnd), and the marker counts the characters actually left out. The old "start + middle ellipsis + end" tests keep their assertions; stewardRequest and every other function in the file stay as they are.

Scope

• core/src/host/steward.js — shorten() only (≈+10/−5).
• core/test/steward.test.js — the new case (≈+17/−0), existing cases unchanged.
• patches/ds-bot/0001-core-fixes.patch — regenerated by node scripts/ds-bot-patches.mjs generate (no hand edit). core/client.js, patches/ds-bot/0008* and the src/client/ mirrors are untouched.
• changelog.d/<this issue>-steward-shorten-chars.fixed.md.

Tier: T2 (core/src/host/**, core/test/**).

Duplicate check

Checked before filing: #1993 (closed) / #2000 (merged, b048e26f509b) fixed five other bounds in this family — steward.js:148 included — and did not touch shorten(); search steward shorten → 0, steward in:title → 6 (all unrelated: parts retry, finish-chunk, condensation ratio), "characters left out of this view" → 0; no open PR touches core/src/host/steward.js.

Done when

The new case is red on the base and green with the fix, core/test/steward.test.js is 11/11, TZ=UTC ARCH_DOCS_BASE=d400edb7 npm run verify is exit 0, and patches check reports the regenerated 0001 series as clean.