feat(engine): language-drift flag on subagent result envelopes (#88)

[rz1989s]

Closes #88 (flag half). The optional resultLanguage system-side preference is explicitly deferred (spec N2) — re-file standalone if wanted.

Spec first (design-first per repo convention): docs/superpowers/specs/2026-09-02-spec-language-drift-flag.md

Summary

glm-5.3-flash drifted into Chinese for all/part of the final report in 4 of 7 dispatches — including two with an explicit English-only instruction. Findings were sound every time; the missing piece was a cheap machine-readable signal. This lands exactly that: flagging, not blocking (the #61 surfacing-only precedent).

Design (D1–D4 in the spec)

• Signal: CJK-family letter ratio (Han/Hiragana/Katakana/Hangul) over the final message — flags at ≥ 30% of letters with a 40-letter floor. Quoted CJK inside an English report stays clean (tested); a majority-CJK body trips it (tested, the observed mixed mode). Latin-script non-English deliberately out of scope (false-positive surface, not the observed drift family).
• Site: finishRun — the single composition point for every terminal path (direct/fallback/turn-budget, pi + claude backends).
• Surface: languageDrift + languageDriftRatio on SpawnResult details; one-line [FLEET] language drift — … warning prefix (composes with the subagent: premature return after a single planning statement (zero work, no turn-budget exhaustion) — re-dispatch succeeds #61 zero-tool prefix); journaled additively on run:ended (the subagent: explicit model string fails where inherited session model succeeds (same model) #59/#49 regression: filesTouched empty on real edit-tool runs — args serialized as empty, run:ended missing the field #60/subagent: premature return after a single planning statement (zero work, no turn-budget exhaustion) — re-dispatch succeeds #61 post-hoc pattern). No renames — frozen surface stays additive-only.
• Deferred (N2): the issue's optional resultLanguage system-prompt preference — it changes every child's substrate and dispatch-prompt placement demonstrably isn't enough; separate decision if wanted.

README

New "Model-quality drift signals" section documents the flag and the #89 controller guidance (reviewer cross-domain references are suspect; verify by git + files, never prose) — #89's armory-fleet-docs half. Its SDD-template half lands in the skills repo separately.

Tests (+11, 835/835)

• Detector ×8: English clean; pure-CJK drift; half-mixed drift (the observed mode); quoted-CJK clean; Hangul; Kana; sub-floor no-flag with honest ratio; empty.
• Wiring ×3: drifted child → prefix + details; English child → no prefix/undefined; journaled run:ended carries flag + ratio (RunLog replay).
• RED verified first (missing module + 2 wiring failures); gates standalone: typecheck ✅, 835/835 ✅.