[hwanjjang]

Summary

On EC2 (2026-11-07, 01:29-03:12 KST) one or two requests of most rounds took 2.4-3.6 s. Each found the newest KMA AWS minute page (nph-aws_txt_min, about 470 KB of EUC-KR HTML) more than 2 minutes old, started the next page and waited for it; the page took 2.1-2.6 s, at the observed 2.5 s bound, so the request usually still answered the previous page. Legacy collected the page every 2 minutes in the background and answered from the stored page.

While a page of the last 20 minutes (the readers' reach) is held, the request now starts the owned acquisition detached and answers with the held page; the next page serves later requests. Without a held page (a cold process with nothing stored) the request waits as before. The cadence (one page per 2 minutes, whatever the outcome) is unchanged.

Changes

• observe.rs ensure_minute: MINUTE_HELD_MS (20 minutes); a held page spawns the owned acquisition and returns nothing to wait for.
• Architecture document: the request-time acquisition table.
• Records: intent/, specs/, plans/issue-2614-kma-minute-page.md; declaration server2/config/tasks/KMA-MINUTE-PAGE.json.

Verification

• New test a_held_minute_page_answers_while_the_next_one_is_acquired: a restarted process with the stored 10:49 page, asked at 10:53 with a provider that takes 3 s, answers within 1.5 s with the held rows, and the next page is acquired. 3 of 3 mutations caught; M1 (the request waits, the behavior before this change) is the Red evidence.
• Full cargo test --locked --workspace 1,214 passed, 0 failed, 7 ignored; cargo fmt --check, cargo clippy -D warnings, placement gate and artifact policy pass.

Decision for AK

• A request that arrives when the next page is due answers the held page, up to about 4 minutes old (2-minute cadence plus the page's own lag), instead of waiting up to 2.5 s for a page it usually missed anyway.

Not in this change: a scheduled fetch without requests, the hourly and city pages, and the other fixes of the EC2 report (#2866 restart reads, #2867 AirKorea refresh).

🤖 Generated with Claude Code

[hwanjjang]

Cross-provider review (Codex gpt-6.1-sol, high): PASS at 7ffae72.

Recommended

• MP-refresh-test (check that a page fetched in the background reaches the next answer): not adopted. The change only stops the request from waiting; the path by which an acquired page is stored, marked dirty and folded for later requests is unchanged by this PR.