Reduce Media Studio payloads and redundant reads

[gateway]

Summary

 Media Studio
- full catalogs and heavy job props at startup
+ route-specific data, bounded workflow summaries, compact preset and job-card props
- repeated page reads and offset rescans
+ scoped in-flight sharing, server-seed reuse and gallery continuations
- Graph media list preloads for restoration
+ exact selected-media reads and bounded fresh picker pages

Preserve existing API defaults, saved-state contracts, pricing, refresh/retry and original-media previews. Add opt-in summary/cursor paths; keep legacy full-list and offset consumers working. Includes template catalog reuse and visible gallery recovery errors.

Evidence

Representative separate paired development comparisons, each with five serial warm samples; these are not one cumulative benchmark:

Scenario	Before	After
Graph startup target reads	11	4
Workflow catalog body across startup + first open	2,162,802 B	14,125 B
Workflow dialog DOM elements	5,078	1,194
Studio decoded document	3,372,797 B	2,780,708 B
Settings decoded RSC	1,215,417 B	90,734 B
Studio preset initial client list reads	2	1
Admin redundant preset reads after server seed	2	0
Gallery backend calls at append offsets 42 / 102 / 498	1 / 2 / 5	1 / 1 / 1
Graph closed-picker decoded metadata	410,614 B	83,274 B
Jobs decoded RSC	2,225,475 B	149,284 B (−93.29%)

• Before: template normalization repeatedly rebuilt catalogs; gallery overlap repeated page payloads. After: isolated contract tests verify once-per-request catalog reuse and freshness; matched gallery traversal eliminates 126,375 redundant decoded bytes while preserving ordered results.
• Before: broad Graph restoration lists omitted older selected media. After: selected previews restore through exact detail reads. Boot HTTP count rises from 3 lists to 12 details; first picker opens remain unchanged and fresh reopens make no metadata reads. This is not a total-image-traffic reduction claim.
• Settings loader reads fall 19→3, but complete capture brackets include independent gallery traffic (19–20→4). Jobs retains all 19 backend reads and 2 development runtime GETs; only card-boundary props change. Cursor responses add 167–168 bytes, while returned backend rows at the measured offsets fall 100/212/548→12 each; SQL scan work is unmeasured.
• Focused checks cover response compatibility/freshness, duplicate and stale page ownership, refreshed seeds, retry/error recovery, dirty-tab and delayed-preview preservation, tied cursor ordering, exact-media restoration/import invalidation, and pricing/prompt projection contracts. Slice-level focused tests and applicable typecheck/lint/file-size/diff checks passed; final Jobs projection adds five pure contract tests. Independent full cumulative Standards and Spec reviews report zero actionable source findings.
• Browser acceptance verifies ordered cards, bounded paging/search, preserved controls/defaults, exact previews, visible failure/retry and original draft/storage restoration. Live conditional pricing/runtime and multi-output Jobs cases were unavailable; contract/source coverage is separate from live proof.
• Memory/canvas and historical Jobs latency investigations produced qualified diagnostics, not additional optimization claims. Retainer/process/GPU, reliable active frames, comparable retention repeats and production coverage remain incomplete. Development timings do not establish causal speedups, leak freedom or a fix for the historical slow sample.

All four hosted CI checks passed on 6059967: publication safeguards, Windows onboarding, assistant mechanical contracts, and the full quality gate. The quality-gate run passed 993 backend tests, 1,120 web tests across 182 files, lint, type checking, the production build and Studio browser smoke. Fixtures ran only in the authorized ephemeral GitHub-hosted environment; no broad local suites or local production build were run. A successful hosted build/smoke is not a production performance benchmark. Private raw captures are intentionally excluded.

Merge Danger

Door: two-way

Source reverts restore prior behavior; no database migrations or destructive data changes. Legacy full-list/timestamp/offset contracts remain available. Roll back coherent frontend/backend cursor or summary slices together.

Blast Radius: cross-stack

Touches Studio gallery/presets, Graph startup/workflow browsing/media selection, General Settings and Jobs card serialization, plus additive backend projections/cursors. Primary risks are stale request ownership, hydration/default preservation, tied pagination order and pricing display fallback. Focused regressions and paired browser checks cover these paths; production and disclosed live-data cases still require follow-up. No merge, deployment or paid-generation approval is implied.