feat(benchmark): show source-bound retained outcome improvement

[loopx-agent]

The six process judgments do not show whether the retained native outcome improved. Add a separate Retained improvement / 已采用成果增量 metric, computed by the existing benchmark-toolkit TypeScript owner from an explicit baseline, native optimization direction and source-bound retention evidence.

Native error 7 → retained 3 displays +4 points. Rejecting a candidate with error 9 while retaining 7 displays 0, with candidate change −2 inside the disclosure. Missing baseline or adoption evidence stays unknown; actual adopted regressions remain negative. This is a provider-declared native comparison, not a composite score, objective completion or LoopX causal attribution.

• Optional comparison extends the existing native-capture contract; omitting it preserves the prior projection shape and model analysis key. An archived submission time may explicitly remain null instead of inventing a clock. Known times retain calendar/window validation.
• Goal overview shows one compact outcome value before the process judgments. Hover/focus/tap explains it; expansion opens baseline, retained measurement and actual retention evidence. Detailed native records move into the existing Sources and usage disclosure. Changed sources remain visibly historical.
• Selecting an analysis scope now creates a browser history entry. Back/Forward restores the real selected scope instead of dropping the previous selection.
• Owning capability: builtin benchmark-toolkit, provider loopx-core. Python remains the existing file/HTTP transport; it only admits the additional exact bound citation ranges. No solver, evaluator, scoring, authority-store, model dispatch default or installed App changes.

Validation

• 17 focused TypeScript assessment/native-outcome tests; 16 Python source-store and production HTTP API tests passed.
• Dashboard + packaged Chat build, control-plane typecheck, documented Ruff and no-argument mypy profiles passed.
• Ego Lite against the packaged bundle and actual local HTTP/native-pack owner: positive/zero/unknown, original retention citation, changed-source historical warning, reload, real selector Back/Forward, hover/keyboard/tap explanation and 390 px overflow checks passed. Synthetic screenshots below; the browser acceptance runner is extended for the same semantics, with its full automated run left to CI.
• Additional file-target mypy trajectory_assessment_store.py probe failed outside the documented profile, reporting 4,920 diagnostics across 582 imported files. It is retained as a failed probe, not treated as a passing check.
• Premerge automatic checks pass; benchmark-sensitive manual review remains required. No merge/install/release is claimed. Private benchmark trajectories and local operating state are excluded.

Remaining integration and CI boundary

This PR supplies the analysis metric and per-Goal renderer. Fixed case-pack registration does not create a real experiment child Goal. The remaining experiment journey must use actual parent/child registration and run/source identities, with comparison at the parent and analysis in the corresponding child.

GitHub kernel-static-checks currently fails the pinned top-level module budget: 147 modules against a limit of 145. Base 3203b55843d2dcca5b77822ed857442adf944587 and this head have the same 147 top-level Python modules; this PR adds none. The hard gate remains held; its budget is not increased to mask the failure. Other CI and exact-head review remain pending.

Visual evidence

Tested source revision: c01d3863cec14cfc21e05724d14212a72a922f2a. Actual packaged browser captures with synthetic input and a disposable deterministic analysis child; these images are not benchmark findings or installed-App evidence. The before renderer uses the same synthetic capture. Desktop is the initial 1716×980 viewport; mobile is the affected research section at 390×844 after scrolling. Native details and optional explanations are collapsed by default.

Before	After

Synthetic local preview (author machine): open packaged research overview.