experiment: read authored behavior paths with explicit test associations - #1121

Problem and value

The existing task-review-map explains changed nodes and accounts for Git hunks, but its generic node list does not say which authored behavior step is associated with which test source. This experiment adds a small sidecar projection over the unchanged, hash-pinned mapping: a missing-artifact path with a test association, a delivered path without declared associations, and a clearly labelled synthetic omission contrast.

Scope was authorized by the maintainer's resumed research task and deduplicated against the existing 49 experiments and open work. Inspired by task-view's Flow view; no Git adapter, upstream renderer or new product schema is copied. Reproduce and inspect.

AI assistance and accountability

• AI: Codex / GPT-6.1-Sol agents for source research, implementation, tests, independent review and actual browser inspection.
• Accountable maintainer: @tt-a1i. Proposed reading improvement over fixed historical public code; flow and test associations are authored statements.
• I reviewed the submitted changes and explanation, understand their impact, and can answer questions about them.

Stability impact

Only experiments/task-flow-tests/ changes. The fixed mapping is reused byte-for-byte from dcedd53856b82d3e403a9651dca0e15840496942. No historical tests or task-view code were executed. The page keeps reference agreement, declared test association and test execution separate. Unknown/duplicate references are rejected. No shared Viewer, CLI, schema, package or installation change.

Scoped test-value/maintenance review retained a small projector and decisive identity, association, allowed-input and output-protection tests. The real-generation benchmark pilot is not applicable: this is a bounded projection/content experiment, not a generation-pipeline change. No comprehension-speed improvement is claimed.

Tests run

Base e837dc4cb83042d2b6c596e039692e25ec62c2da; implementation 0d5f46834bbe8e631b6c20d2acc81c206583812d; head 843c877b251baf89925f99e11594fc5630c1c165 adds browser evidence only.

• 9 focused tests pass, including a single-path/zero-association case that exposed and now prevents an incorrect fixed summary.
• Finalize: validate/deliver/check pass with 0 diagnostics; default Explainer browser-check not requested.
• Separate strict-provenance visual-check: containment/readability/chrome/theme/captures pass, but overall fails with 16 keyboard-traversal diagnostics across four states. Exact failure, initial and revised drafts, and an intermediate test failure remain recorded.
• Root actual browser: step navigation, regression source link, light/dark desktop and 390px viewport inspected. Actual Tab traversal reached all three footer source links and the Archify link at steps 44–47. This is separate evidence; the automated failure is not relabelled as passed. No assistive-technology claim.
• Independent review passed the final semantic/identity/maintenance scope. Remote CI is recorded separately in the dated follow-up below.

Visual evidence

Final artifact SHA256 f8c908b1d9e4e2eccf127387d886eba425aa06106448fbd212eac931f6f4b17b. Actual screenshots at 1440×900; the 390px image and full keyboard sequence are in root-review. Agent visual inspection confirms the linked-test state is visible; the mobile diagram is a small preview with an explicit open-diagram affordance and long code scrolls locally. Human acceptance remains pending.

Generated artifacts

Only isolated experiment HTML, JSON, screenshots and portable evidence summaries. Shared Viewer/ZIP inputs are unchanged. Original and second-draft outputs preserve their own evidence identities. Source and license attribution distinguish task-view's MIT root from vendored EPL-2.0 materials; no vendor code is reused.

CI follow-up · 2026-10-27

精确 head 的初次 browser-regression (2) 失败于共享 Sequence 的 1396×480 短视口 containment。第一次单 job 重跑请求因 workflow 尚未结束而返回 job cannot be rerun。随后 webm-decode 成功结束；webm-artifact 当次失败来自 browser 依赖汇总，没有第二个解码错误。

workflow 完成后，单独请求重跑该 browser job。相同 head 的 attempt 2 于 00:31:02 UTC 成功结束；当前检查为 19 success、2 deployment skipped。仅失败 browser job 及其依赖汇总重新执行，其他成功 job 保留原执行时间。

原始失败和本地 Reader Layout 18/18 通过仍保留；重跑通过没有确定首次失败根因。未修改共享代码或启动全套新 run。首次只读诊断 · 重跑记录 · 最终检查快照。

Existing keyboard candidate validation · 2026-10-27

复用既有远程候选 codex/lab-keyboard-traversal@c96e8aa7，对本 PR 原始最终 HTML 直接运行其现有 browser-check API，四种视口/主题状态均完成 40/40，47 步，零诊断，HTML 前后字节一致。没有创建重复修复分支，也没有修改本 PR 产物。

这是旧候选对冻结输入的新验证，不代表当前 dev 已集成、完整 CLI 集成已验证或原失败 receipt 已改成通过。范围、复现入口与原件摘要。

上述候选随后刷新到当前 dev，并提交为 Draft #1124。它尚未合入 dev；本 PR 的历史 HTML 和失败回执保持原字节。当前 CI 结果见上方 attempt 2 记录。