feat(research): time-anchored Explore history lens for late-phase attribution

[huangruiteng]

Stacks on #6616 (does not replace its reader). Validated on both accepted datasets: the PR's redacted EdgeBench capture (9.6 days, 364 events) and a real local long game research Goal (2.9 days, 483 events, 51 path_delta decisions), through the packaged Chat over the real loopback history owner.

Why

Research collaborators and product users need to answer, without scanning a node grid:

• 长程中后期是否有进展 — equal thirds of record time show new questions, source finding labels, explicit decisions and recording gaps.
• 为什么有显著进展 — findings are split into the writeback owner's Input revision / Applicability / Observation / Interpretation sections; path_delta decisions expose Outcome plus retained / changed / stopped.
• 为什么无进展 — confirmed-without-adoption, tentative and refuted stay distinct; unstructured findings explicitly say decisions cannot be separated instead of inventing them; recording gaps are visible as gaps, not inactivity.

What

• New default Timeline lens in the existing Goal Overview 研究历程 section; the old graph stays one click away as 关系图, with the same selection, focus, refresh and pagination code.
• Record-density histogram with gap bins; routes are connected components of recorded typed edges and parent grouping only; unlinked questions render as a beeswarm. Time adjacency never creates an edge; marker position never encodes progress.
• One optional click opens the existing detail panel, now showing structured observation / decision evidence and version/applicability provenance. Unstructured findings remain raw prose.
• Phase card click filters markers and the decision trail to one third; a "whole history" control returns. Thirds are labelled as reading groups, not a pre-declared evaluation window.
• New shared typed presentation module (loopx/control_plane/presentation/explore_timeline.ts): section parser, thirds, density bins, routes via edge/parent union-find, decision trail. Display-only — no protocol vocabulary, graph authority, evaluator or progress score.

Validation

• node --test tests/control_plane_ts/explore_history.test.ts: 6/6 pass (added thirds/routes/decision parsing case incl. "unstructured stays raw").
• pytest tests/presentation/test_explore_history.py passes; dashboard tsc --noEmit, CSS custom-property check and packaged vite build + build:chat pass.
• Packaged browser pass with both datasets: full pagination, marker/detail selection, phase filter, 390px zero horizontal overflow with reduced motion.
• Semantic inventory advisory: no new Enum/Literal vocabulary carriers.

First-screen screenshots were reviewed locally against both datasets before this PR; no private screenshots are uploaded. The companion Lark baseline experiment used the redacted EdgeBench capture only.

Not included (separate owning work under #6606): native outcome curves, agent deep-analysis action, Lark sink changes.

[huangruiteng]

Added the reusable Lark-whiteboard renderer (examples/research-history-lark/board.mts) using the same shared timeline model; output verified identical to the reviewed prototype on both datasets.

[loopx-agent]

Product integration assessment at 916cec603e8537c0791a9bb4042e07538c579e60: the record-time lens is a useful addition to #6616, especially for long histories, recording gaps and explicit decisions. It remains a display projection, not a progress score or causal evaluation. An unchanged merge is premature.

The integration reproduces and repairs three reading failures: phase filtering after per-question deduplication hides an earlier decision when that question has a later decision; non-replan tick marks label wait/stop/no-change as continue; append order can make a delayed historical backfill look like the latest recorded decision. The packaged markers also inherit a 44px minimum height while keeping a 14px width, stretching points into ovals. The refined candidate retains all decisions, selects phases in the shared TypeScript presentation owner, preserves the actual outcome and separates the circular glyph from its touch target.

The consolidated proposal also carries #6616’s later UTC-day-preserving capture and frozen snapshot identity correction. It discloses the static Lark example’s excerpt limits and reuses the same decision selector. The complete event-history reader and a compact current-state projection remain different sources; missing native edges, attempt intervals, costs and adoption bindings remain unknown.

Current author validation: the real scoped HTTP history test, 24 TypeScript presentation/digest regressions, dashboard typecheck, packaged build/integrity and semantic vocabulary check pass. Packaged research-history, ordinary Goal showcase, managed results, research results and blocked-reason journeys pass, including failed-refresh recovery. These do not establish independent review, installation, full V11 acceptance or a completed V12 analysis service. Independent exact-head review has been requested; the consolidated first-screen preview is awaiting owner confirmation before publication. This integration note is not an approval verdict.