[codex] Upgrade Agent Studio workbench

[ColinLi98]

What changed

• Upgrades Agent Studio into the interactive local workbench: startup page, reader-first Studio layout, director panel, choice cards, branch map, quality/export status, and .nosbook export support.
• Adds ReaderShell async generation hardening for queued /v1/reader/continue responses with job polling, resume-on-stale, session reload, and visible wait state.
• Adds Agent Studio rendered smoke coverage, desktop/mobile screenshot artifacts, sticky director checks, mobile bounded choice scroll checks, and visual review checklist output.
• Adds scripts/run_agent_studio_local.sh, which starts the local backend and automatically opens /app?product=author&workspace=studio&debug=1 after /health passes. Set AGENT_STUDIO_OPEN_BROWSER=0 to disable browser opening.
• Refreshes the standard cross-pack benchmark baseline after CI showed the initial-import baseline was stale while current benchmark quality passes at 1.000.

Merge gate evidence

• Lane: Lane B
• Phase: Phase 2
• Task: Task AS-4/AS-5 follow-up hardening
• Goal met: yes
• Out-of-scope changes introduced: no
• Tests run: targeted Agent Studio/ReaderShell/frontend smoke contract tests, targeted AuthorWork branch/export tests, cross-pack merge gate tests, and local rendered Agent Studio smoke
• Benchmark / eval run: standard cross-pack benchmark plus merge gate after baseline refresh
• strongest pack delta: current strongest remains synthetic_min_pack and urban_mystery_lotus_lane after baseline refresh
• weakest pack delta: current weakest remains jade_court_romance, tide_archive_memory_debt, and xianxia_forgotten_vow after baseline refresh
• cross-pack pass-rate delta: +0.000 against refreshed baseline; prior stale baseline showed +0.067 before refresh
• issue category delta (Q03/Q04/Q05/Q09 if relevant): no generation/planner changes; benchmark phase gate reports no blocking issue-category regression
• rollback point: revert the PR branch commits, especially the Agent Studio UI/smoke changes and refreshed tests/benchmark_baseline.json
• next suggested task: run long-route benchmark evidence after Agent Studio PR lands

Product impact

• Does this move commercialization forward?: yes, it makes Author-side local creation usable and CI-verifiable.
• Does this improve kernel/product/ops instead of just current-pack polish?: yes, changes are product shell, smoke reliability, PR review process, and benchmark evidence hygiene.
• Does this make weakest packs easier to diagnose or improve?: yes, cross-pack benchmark and visual smoke artifacts now surface current weakest/strongest packs and UI regressions clearly.

Validation

• bash -n scripts/run_agent_studio_local.sh scripts/run_agent_studio_smoke.sh scripts/run_frontend_shell_smoke.sh scripts/run_reader_shell_smoke.sh
• node -c src/narrativeos/web/agent_studio.js && node -c src/narrativeos/web/reader_shell_v2.js && node -c scripts/verify_agent_studio_smoke.js && node -c scripts/verify_frontend_shell_smoke.js
• .venv/bin/python -m py_compile scripts/write_agent_studio_smoke_step_summary.py scripts/write_frontend_shell_smoke_step_summary.py
• .venv/bin/python -m pytest tests/test_agent_studio_interactive_workbench.py -q
• .venv/bin/python -m pytest tests/test_frontend_shell_smoke_ci.py::test_frontend_shell_smoke_scripts_exist_and_are_parseable tests/test_frontend_shell_smoke_ci.py::test_agent_studio_smoke_workflow_wires_headless_runner_and_artifacts tests/test_frontend_shell_docs.py -q
• .venv/bin/python -m pytest tests/test_reader_shell_v2.py -q
• .venv/bin/python -m pytest tests/test_reader_shell_flow.py -q
• .venv/bin/python -m pytest tests/test_author_works.py::test_author_work_flow_supports_generate_edit_diagnostics_and_submit tests/test_author_works.py::test_author_work_can_create_parallel_universe_branch_without_overwriting_mainline tests/test_author_works.py::test_author_work_branch_discards_mainline_future_chapters_after_selected_fork_point -q
• /tmp/narrativeos-py312-venv/bin/python -m pytest tests/test_cross_pack_merge_gate.py tests/test_cross_pack_benchmark.py::test_cross_pack_benchmark_outputs_kernel_metrics tests/test_phase0_guardrails.py -q
• /tmp/narrativeos-py312-venv/bin/python -m src.narrativeos.benchmark.runner --baseline-file tests/benchmark_baseline.json --database-url sqlite:///narrativeos_beta.db --markdown-out /tmp/pr3-benchmark-summary-updated.md > /tmp/pr3-benchmark-updated.json
• /tmp/narrativeos-py312-venv/bin/python -m src.narrativeos.benchmark.merge_gate --benchmark-file /tmp/pr3-benchmark-updated.json --summary-out /tmp/pr3-merge-gate-summary-updated.md
• CI_HEADLESS=1 APP_PORT=8018 CHROME_PORT=9238 CHROME_USER_DIR=/tmp/narrativeos-chrome-agent-studio scripts/run_agent_studio_smoke.sh
• APP_PORT=8766 AGENT_STUDIO_OPEN_BROWSER=0 bash scripts/run_agent_studio_local.sh against a temporary /health stub, confirming the script emits the Studio URL.

Notes

Full tests/test_author_works.py was attempted but did not finish locally and was stopped after repeated high-CPU runs; the Agent Studio-relevant AuthorWork subset above passed.

[ColinLi98]

Agent Studio visual review accepted from artifacts/agent_studio_smoke_visual_review.md in the latest PR run (25705903844).

Viewport	Check	Status	Evidence	Reviewer note
desktop	Three-column workbench review	manual_review	artifacts/agent_studio_smoke_desktop.png	accepted
mobile	Stacked workbench review	manual_review	artifacts/agent_studio_smoke_mobile.png	accepted