[cocolord]

Goal And Delivered Outcome

The synthetic delegation Host inherits a locale decoder on Windows CPython 3.11 even though its existing parent transport sends UTF-8 JSON. Unicode input can therefore fail the Host before the independent artifact validator runs, leaving the original-task recovery cases without task_failure.

Configure only the fixture Host's stdin as UTF-8. A File/SQLite regression simulates CP1252 in that child, rejects an invalid artifact, repairs it and revalidates the cached result with the original turn key and exactly one Host invocation. Related to #8360; target main, author baseline 4e949f405528fd8087f0623526c62f890c1ce982.

Author Declaration

• Written by: model_agent — TRAE / GPT-6.
• Specification: docs/architecture/rfcs/monorepo-distribution-split-v0.md §9 V4 and the existing generic Host transport / original-task revalidation contracts at the author baseline.

Criterion	Disposition	Path	Evidence
Consume the existing UTF-8 pipe protocol	implemented	tests/test_local_delegation.py::HOST	Same-environment immutable baseline Host fails the new oracle; corrected Host passes.
Preserve original task identity and avoid a second Host execution during artifact repair	implemented	New File/SQLite regression and tests/test_delegation_task_revalidation.py	Source and isolated-wheel recovery checks below.
Preserve failure and remaining acceptance boundaries	implemented	Paired RFC non-normative checkpoint	Original Windows and Frontstage failures retained; no performance or full-kernel claim.

• Self-check: inspected the actual transport and validation owners, baseline sensitivity, complete recovery suites, package provenance, paired translation and public boundary. The adjacent refactoring pass found no additional extraction warranted: the encoding fix belongs to the existing source-owned Host fixture. No production rule, persisted state, protocol vocabulary or default Host environment changes.

Scope And Continuation

Complete fixture correction with a durable regression and paired qualification record. The parent minimal-kernel program stays open for complete provider-first transaction, PostgreSQL host/wheel, complete costs and adoption acceptance. Retain I3 for at least one minor release and all historical/frozen failures. #8360 retains its own exact-head review/CI gates. Native Windows verification of this new head remains pending; the prior failed run is retained. This proposal does not merge or upgrade an installed Host.

Validation

• Tested fixture commit: a4ce613585fdb329ba3600f0ede768864214ed84; final paired-doc head: 107c4fab80406a96bb0f6ccfd62b47230034d2fc.
• Run state: finished locally; CI pending after submission.
• Input classes: synthetic, public_fixture.

Check kind	Result	Public-safe evidence / limitation
regression_parity	passed	CPython 3.13.5 / Node 22.22.3 on macOS arm64: replacing only the corrected fixture HOST with the immutable baseline makes the Unicode task-validation oracle fail. The corrected File/SQLite cases both pass. Earlier invalid probe preparations are retained separately.
integration / real_entrypoint	passed	tests/test_local_delegation.py plus tests/test_delegation_task_revalidation.py, CPython 3.13.5 / Node 22.22.3, 81 cases pass using temporary File/SQLite stores and synthetic subprocess Hosts.
integration / real_entrypoint	passed	CPython 3.11.15 / Node 22.22.3 source: new two provider cases plus all 12 existing original-task recovery cases pass, including MCP disconnection and lock-timeout recovery.
integration / real_entrypoint	passed	Fresh isolated monolithic wheel, CPython 3.11.15 / Node 22.22.3: the same 14 cases pass. Wheel imported before source-owned fixtures; all 1,920 present packaged tracked runtime entries match the source. The same 17 source-only exclusions remain: canary baseline, host plugin/config, fault-test helper and documentation/media.
static	passed	Ruff and paired staged RFC check; diff advisory introduces no shared vocabulary. Public-boundary subcheck passes for the three candidate files. Full local loopx check still fails on unrelated registered-Goal authority errors; private diagnostic contents are excluded.
static / integration	passed	Standard diff-derived loopx canary premerge --from-git-diff --git-diff-base origin/main: all five direct and 11 selected checks complete, zero blocking failures. The unchanged module-metric ceiling finding in loopx/extensions/lark/private_conversations.py is reported as an inherited advisory. The first attempt failed because this new worktree lacked TypeScript dev dependencies; npm ci --ignore-scripts prepared them, and the unchanged-head rerun passes. No gate was relaxed.
integration	not_run	New exact-head native Windows CI pending. The original #8360 Windows arm has six recovery failures. Its independently retried Frontstage manifest request still returns HTTP 504. #8360 also retains separate digest-declaration and selected-owner loading failures; this fixture correction does not resolve those gates.

The changed fixture is exercised through the real generic Host subprocess transport and real temporary File/SQLite providers. No model or real-agent execution, PostgreSQL host/wheel qualification, performance campaign, split installation or installed adoption is claimed. Existing recovery and failure classification remain intact.

Frontend / Visual Evidence

UI impact: none. No UI or documentation chrome changes; screenshots and preview are not applicable.

LoopX Area

• Host or runtime integration
• Public docs or presentation surface (RFC evidence only)

Shared-authority RFC fixture impact

N/A: no TypeScript migration or shared-authority production change.

Boundary Checklist

• Public-safe diff/body; no private state, credentials, raw traces, internal links or local paths.
• Scoped fixture repair; no competing implementation or benchmark work.
• Both English/Chinese RFC files change together.
• UI impact marked none.
• Every commit includes a DCO sign-off.