fix(deep-interview): canonicalize Hangul before hashing and capping prose

[yazzang-homelab]

What

Canonicalize deep-interview prose to NFC at the identity and length boundaries: questionHash(), answerHash(), buildAnswerShell() (stored question_text / selected_options / custom_input), and assertDeepInterviewInputWithinLimit().

Why

Fixes the first half of #3871.

Hangul reaches the interview in both composed (NFC) and decomposed (NFD) form — macOS-sourced pastes and several clipboard/IME paths emit NFD — and the two forms are one text for the user but two different JavaScript strings for us:

19 39 false false   # [...nfc].length, [...nfd].length, questionHash equal?, answerHash equal?

That produced three Korean-only failures:

• appendOrMergeRound() saw a different answer_hash for the same answer, so its documented noop became a duplicate/replaced round — one answer, two round histories.
• assertDeepInterviewIntentReview() matches approval evidence by answer_hash against recorded answers, so crossing a normalization boundary invalidated the user's own approval.
• Caps are code-point counted, so decomposed Hangul cost 2–3 units per syllable: MAX_USER_RESPONSE_LENGTH = 10_000 rejected Korean answers at roughly a third of the advertised budget while ASCII prose got the full 10,000.

#3716 asked models for literal UTF-8 in tool inputs, but it is a prompt directive; nothing in the runtime canonicalized user prose. This is the code-level half.

Testing

packages/coding-agent/test/deep-interview-hangul-normalization.test.ts — 5 tests: fixtures are genuinely distinct strings, one question identity across forms, one answer identity for both selected options and custom input, appendOrMergeRound returns noop with a single round when both forms arrive, and cap parity (10,000 composed and decomposed both pass; 10,001 decomposed throws).

bun test packages/coding-agent/test/deep-interview-hangul-normalization.test.ts   # 5 pass (4 fail on dev without the src change)
bun test packages/coding-agent/test/deep-interview-workflow-gates.test.ts \
         packages/coding-agent/test/deep-interview-skill-contract.test.ts \
         packages/coding-agent/test/deep-interview-gate-redteam.test.ts \
         packages/coding-agent/test/sdk-workflow-gate-emitter.test.ts \
         packages/coding-agent/test/agent-session-deep-interview-continuation.test.ts   # 102 pass
bun --cwd=packages/coding-agent run check   # biome + tsc clean

Also verified as not the cause and left untouched: replaying all 173 Korean ask questions from the local session corpus through renderDeepInterviewAskQuestion() / formatDeepInterviewSelectorPrompt() at widths 40–100 loses no Hangul and drifts no padding, so #1996 holds.

GJC verdict

gajae.pr-review-verdict.v1 needs-human sha256:a366111caf273ff2761e9267476ecaa157c8aae8 reviewer:human evidence:rebased onto dev 6c1c8a4798917be42071af02ab3c005ca71817c3 (exact-head CI base gate) + bun test packages/coding-agent/test/*deep-interview*.test.ts (89 pass) + bun test packages/coding-agent/test/gjc-runtime/ (1259 pass) + bun test tools/ask, workflow-gate-redteam, workflow-mutation-guard, workflow-state-command, gjc-skill-state-hooks (189 pass) + tsc -p packages/coding-agent --noEmit clean + biome clean on touched files

---

• Target branch is dev
• bun check passes (bun --cwd=packages/coding-agent run check)
• Tested locally
• CHANGELOG updated (if user-facing)
• Verdict above matches the exact PR head, not an earlier commit

[yazzang-homelab]

Rebased onto current dev (11e48d5bc, #3901) — the previous head was CONFLICTING/DIRTY on the CHANGELOG ## [Unreleased] section. New head 392401cf6; rebase was CHANGELOG-only, no code merges, and the diff is otherwise unchanged.

Re-verified on the new base: bun test packages/coding-agent/test/deep-interview-hangul-normalization.test.ts → 5 pass, 0 fail.

[Yeachan-Heo]

This item is being moved to the Tier 3 review lane. Further review is deferred until after the current emergency stabilization period.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]