[Performance][P1] Validate responsive covers and resilient IPFS delivery

[knzeng-e]

Outcome

Validate the delivered responsive-cover pipeline under live upload, mobile, and adverse gateway conditions.

Delivered

• Server-generated responsive WebP variants and intrinsic metadata.
• Responsive image selection, fixed aspect ratios, placeholders, lazy loading, and active-artwork priority.
• Bounded gateway fallback behavior and browser regression coverage.
• Original cover retained as the archival asset.

Remaining acceptance

• Publish a representative live cover through the production API and verify the generated variants.
• Measure mobile transfer and paint behavior against the existing budgets.
• Exercise slow, failed, and CORS-rejected gateways with live or controlled endpoints.
• Confirm immutable CID cache behavior and decide whether a read-through edge is justified.
• Define any legacy-cover backfill from measured need.

This validation should run with W13/#88 rather than reopening the delivered image architecture.

Local scope: docs/backlog/24-access-streaming-v2.md.

[knzeng-e]

UX audit evidence (2026-09-17)

Complement from docs/design/ux-design-audit-2026-09-17.md (finding UX-01). No scope change; this issue keeps responsive variants, gateway racing and caching.

Measured from a local macOS network against the live catalog (9 releases, 7 distinct cover CIDs), one sample per cover:

Gateway	Result
gateway.pinata.cloud	HTTP 200 for all 7 covers, 4.9 s to 6.7 s each; originals from 61 KB to 2.2 MB (two covers above 2 MB: a 2.2 MB PNG and a 2.1 MB JPEG)
paseo-ipfs.polkadot.io	No response from this environment (curl status 000); not diagnosed further

Rendered effect in Chromium at 1440 px and 390 px: the 1.2 s placeholder budget from W08 fires on almost every cover, so a first visit shows a row of near-identical generated placeholders; covers then paint progressively over them. The placeholder-text and swap presentation are handled in #176 (W21); the variant pipeline remains here.

These are single samples from one network, not p75 evidence.