[cj2026-bit]

Problem

Chat source inspection stalls when repeated retrievals accumulate hundreds of results. Metadata layout measurements, eager detail rendering and repeated source derivation add work, while the panel mounts every card. AIDP catalog pages default to 10 records and leave incomplete four-column rows.

Fix

Backport the final changes from #4213 onto the v2.8.1 tag (30c5168452440dcf66c6785e1f478f9d2b4d3f4d) for main.

Head: 513cb6264e38ae45f0f03fb2566142c3619ec62c. Both source and target branches are in ModelEngine-Group/nexent.

• Clip metadata summaries to two rows and render nested details/JSON on demand, preserving hover and pinned inspection.
• Share message source snapshots, isolate panel actions, memoize cards and paginate sources/images at 20 per page. Both inline citations and the results summary remain usable; actual citation identities locate the selected page.
• Deduplicate uncited historical images, avoid placeholder-index React keys, and isolate downloads/errors by message while preserving same-message card state.
• Coalesce foreground pin/activity refreshes within the five-second freshness window. Preserve the v2.8.1 Workbench-only activity query/run-end scope; do not import develop's unrelated expansion to Agent-chat global polling.
• Default the AIDP catalog to 12 with 12/24/48 options. Card/table switching retains the page, filters and records without a new list request.
• Adapt existing tests only and consolidate the complete scope into one change record. No new frontend UT, Case IDs, Feature changes or performance helpers.

Verification

• All seven current GitHub checks passed for the above head, including type checking, amd64/arm64 web builds and CodeQL.
• Executed on this hotfix checkout: 26 selected existing frontend UT passed; the existing source identity and pagination Case scripts passed all 21 checks; AIDP catalog checks passed 6/7. The remaining unchanged deletion-modal test has the previously reproduced baseline failure (cancel button not found), so the full selected run is 53 passed / 1 failed, not a fully passing suite.
• TypeScript, targeted ESLint (zero errors; five existing warnings), changed frontend format, copyright and whitespace checks passed.
• Formal asset design validation passed. Unrelated baseline-generated Feature navigation drift was excluded from this PR.
• Full implementation asset validation is blocked by a pre-existing v2.8.1 record: CHAT-AGENT-DETAIL-20261026 references the absent frontend/tests/workbench/ConversationResumeBranch.test.tsx. This hotfix does not add a UT or change that unrelated record to suppress the error.
• Source files were ported without unrelated develop commits. At preparation time, main differed from the tag only by its subsequent VERSION update; this branch does not change VERSION.
• Deployed this hotfix SHA to the local DEV deploy checkout and exercised a new persistent AIDP-shaped 300-source conversation. Both entry points, all 15 pages (300 unique keys), boundary/last/repeated citation selection, hover and full large-JSON equality passed. Five source open/close cycles added no API requests. Frontend and config/runtime/mock services remain running; data is retained.

Warm foreground measurements on the deployed hotfix:

Action	Samples	P50	P95	Maximum
Open results summary	20	230.4 ms	291.3 ms	297.2 ms
Open inline citation	20	217.8 ms	249.2 ms	353.1 ms
Switch between first/last source pages by citation	20	233.0 ms	283.6 ms	308.7 ms
Hover metadata, including its existing delay	10	232.6 ms	308.8 ms	308.8 ms

These are input-to-two-animation-frame approximations in development mode, not production SLOs. Earlier near-integral-second scheduling samples are retained separately. Foreground observation still includes long tasks (maximum 301 ms); the measurements do not establish that every page performance issue is resolved.