Filed by the Account-2 read-only amplifier (2026-09-16, HEAD f635e62). Problem: on the readable/EPUB path, all pages resolve before the first asset request starts: pageEntries = await mapLimit(spine, 4, …) completes fully (src/archive-read.mjs:114-127), then the dedup pass runs (:130-131), then assets begin (src/archive-read.mjs:137-152). A magazine pays N page-RTTs then M asset-RTTs back-to-back, although each page's refs are known the moment that page lands. Coordinator-verified against the tree.
Candidate shape: feed each page's assetRefs into a shared deduped Set+queue as pages complete, so asset RTTs overlap the page tail; preserve first-seen ordering and the existing fail-fast semantics (mapLimit/pool).
Impact: for magazine-shaped titles (large M, modest N) this could hide a meaningful fraction of the asset phase under page downloads — same shape of win as #18's entry-time prelaunch. Unmeasured; med confidence.
Evaluation question (implementation is NOT the todo until measured): is the added concurrency complexity worth it for typical M/N shapes in sample-data? Measure first in perf/archive-parallel-sim.mjs (same scenarios, before/after); adopt only on a stable ≥10% win, consistent with the recorded adoption bar (#19).
Proposed focus packet: shared asset queue fed from the page worker; harness A/B; watch ordering of fetching asset i/M logs and staging-write contention.
Filed by the Account-2 read-only amplifier (2026-09-16, HEAD f635e62). Problem: on the readable/EPUB path, all pages resolve before the first asset request starts:
pageEntries = await mapLimit(spine, 4, …)completes fully (src/archive-read.mjs:114-127), then the dedup pass runs (:130-131), then assets begin (src/archive-read.mjs:137-152). A magazine pays N page-RTTs then M asset-RTTs back-to-back, although each page'srefsare known the moment that page lands. Coordinator-verified against the tree.Candidate shape: feed each page's assetRefs into a shared deduped Set+queue as pages complete, so asset RTTs overlap the page tail; preserve first-seen ordering and the existing fail-fast semantics (mapLimit/pool).
Impact: for magazine-shaped titles (large M, modest N) this could hide a meaningful fraction of the asset phase under page downloads — same shape of win as #18's entry-time prelaunch. Unmeasured; med confidence.
Evaluation question (implementation is NOT the todo until measured): is the added concurrency complexity worth it for typical M/N shapes in sample-data? Measure first in perf/archive-parallel-sim.mjs (same scenarios, before/after); adopt only on a stable ≥10% win, consistent with the recorded adoption bar (#19).
Proposed focus packet: shared asset queue fed from the page worker; harness A/B; watch ordering of
fetching asset i/Mlogs and staging-write contention.