[pelikhan]

Result

Optimizes only Model routing run detail (routing-run), not the overview. Five alternating pairs on the same fully ingested deployed dataset reduced populated-run median readiness 10.438 s → 0.464 s (95.6% lower, 22.5× faster). Readiness waits for all five views and worker responses, cleared pending requests, and settled rendering—not just the shell.

Declarative change and necessary generic support

Routing selection, latest-attempt windows, accounting, correlation, ordering, route bindings and bounds remain in Dashboard Language JSON. The JSON now preserves routing-run-key in correlated accounting, joins on run key + event, and explicitly bounds the one-row accounting view.

The generic compiler honors those declarations by propagating constraints only through preserved window partition keys and mandatory equality join keys. Computed/shadowed fields, limits, nonpartition windows and cross-run baselines remain barriers. The generic native planner evaluates authored run-derived expressions on canonical run identities, then reads evidence through the existing byRun index. It does not parse/invert route keys, recognize routing IDs, create indexes or add a cross-view cache. Matching null-extended orphan identities force fallback.

Remove the app's imperative source-query safety-cap injection and obsolete rewrite helper. Its unrequested limit:100000 ran before route scope and blocked legal pushdown. Authored root/dependency limits retain their semantics. Canonical pagination and fail-closed output budgets remain authoritative; oversized selected scopes still fail rather than truncate. Browser-capacity diagnostics remain.

This complements #19005's exact raw tuple-key seek and #19007's shared page/DAG execution. Neither mechanism is reimplemented; their shared-engine patches apply cleanly. Regenerate catalog JSON from composed declarations when combining branches.

Repeated paired measurements

Baseline checkout: 2de03249512f65fc0eb7ca2a438eac647205f2e2. Both variants use the same hash-verified deployed snapshot, manifest SHA-256 cce19a1f7237786b47a363678ea51e6a685e93a2eca21dd1bd80aac82b605808. Retention pinned to 2026-10-19T18:09:55Z: 575,582 records, 4,073 runs, 138,032 audits throughout. Each variant opens a fresh page against the fully ingested database; alternating pair order accounts for concurrent shared-host work. These are local paired results, not a direct comparison to the historical public 8.77 s measurement.

Populated run: githubnext/gh-aw-cao:39042396467.

Pair	Order	Before	After
1	Before / after	9.931 s	0.475 s
2	After / before	10.299 s	0.464 s
3	Before / after	10.438 s	0.409 s
4	After / before	11.587 s	0.446 s
5	Before / after	12.029 s	0.551 s

Representative	Median before	Median after	Before range	After range
Populated	10.438 s	0.464 s	9.931–12.029 s	0.409–0.551 s
Retained evidence-free run (github/gh-aw:37668208348)	10.478 s	0.413 s	9.960–10.514 s	0.379–0.520 s

All 50 view comparisons matched rows/order and availability/completeness/freshness/as-of metadata. Populated outputs remain 1 / 0 / 2 / 32 / 0 rows; the evidence-free run remains empty in all five views. No browser errors.

A separate untimed production-worker storage diagnostic confirms each detail view reads exactly one audits.byRun partition of 99 records, not all 138,032 audits. Identity reads remain: two 4,073-run reads and one 695-workflow read per view. Native reads are omitted by the existing debug recordsRead counter; zero there is not claimed as zero actual work.

Correctness and validation

Regressions cover production-worker parity, exact/correlated ownership, complete/partial/unknown/zero costs, phase separation, resolution roles, missing/malformed/filtered runs, latest-attempt refresh, orphan fallback, cross-run baselines, authored limits and one indexed Audit read per view. Browser coverage checks filtering, missing-run/back navigation on desktop/mobile with IndexedDB and SQLite.

Passed: 371 dashboard unit suites (3,600 tests and one skipped before deleting obsolete cap-helper expectations), final cleanup-focused suites (30 tests), dashboard lint/document validation/typecheck, eight routing browser tests, Overview worker-memory gate, 71 workflow compilations, file-size/diff checks and 36 generated catalog contract tests.

Limitation: the general Lighthouse run scored 99–100 on empty scenarios, but populated-data ingestion timed out after 900 seconds. The complete Lighthouse suite is not claimed as passing. The routing-specific paired benchmark above completed against all original canonical records.

Populated desktop/mobile screenshots will be attached in a PR comment; no screenshots are committed.

[pelikhan]

Populated full-site Model routing run detail after all five views settled, using the pinned deployed dataset. Desktop (1440 px) and mobile (390 px); screenshots are not committed.

[github-actions]

Dashboard dead views

Dead element views (0):
Dead dashboard pages (0):
Analyzed 8 element renderers and 66 dashboard pages across 1 dashboard.json file(s) from dashboard/site/src/components/ui-elements.js.

⚠️ This reachability report is informational and does not block the pull request.
🔎 View the workflow run and report artifact.

[pelikhan]

Superseded by the single consolidated draft PR #19015 at f0b769c. The reviewed source changes and preservation corrections are collected there, with regenerated catalogs and combined validation. Closing without merging or deleting the source branch; history and original evidence remain available.