TODO-001: todo board IO — trust readdir, reuse known paths

Now: DONE

The board render does ~4.4k syscalls for one screen of one-liners: every key is re-probed through pageFile's 6-candidate stat ladder though listTopic just got the exact name from io.readdir, and work.js re-discovers the board + re-probes the same pages once per wt row. NARROWED (gritzko 2026-08-03): this ticket is path reuse ONLY — derive the page file from the name already in hand, memoize per run; read-size work (bounded head / single read) is out of scope.

Input

Context

Profiled jab todo 2026-07-24 (786 tickets on disk, warm cache): 4403 syscalls — 888 opens, 854 stat + 757 lstat, 1535 reads.

views/todo/todo.js listTopic: readdir yields KEY.mkd / KEY/, then pageFile(dir, key) re-probes up to 6 paths (3 exts thin + 3 fat) — ~800 avoidable stats per board.
readBytes reads min(st.size, 1 MiB) — the WHOLE file — but pageTitle uses bytes up to the first \n only.
views/work/work.js is a second payer: ticketTitle/ticketLink per wt row call todo.boardDir() (a stat) + pageFile probes + a full pageTitle read, 93 rows, zero memoization.

Goals

No path is probed twice: a name readdir returned, or a file already found this run, is never re-discovered.

listTopic derives the page file straight from the readdir entry (KEY.<ext> file, or KEY/ dir → README probe only then); pageFile stays for direct todo KEY addressing.
work.js memoizes boardDir + per-key title/link per render run (or todo.js exposes a memoizing lookup for consumers).
strace acceptance: the ~~800 pageFile re-probe stats are gone from jab todo; jab work's todo-side stats drop~~ 3x.
OUT OF SCOPE (narrowed): pageTitle read size — the full-read-serves-/todo/TODO/TODO-002 plan moves with that ticket.

Constraints

Rendering stays byte-identical (existing todo/work tests are the harness); todo KEY direct addressing keeps working for closed tickets and fat layouts.
No caching across runs in this ticket — the wt-level title index is a separate discussion (deliberately out of scope here).
The 1 MiB CAP stays for emitPage (full page render still reads the file).

WIP

Design decisions

TODOs

repro: syscall-count harness (strace or io-op counter) pinning the current cost, then the reduced cost.
listTopic: filename from readdir, no pageFile re-probe.
work.js per-run memo for boardDir/ticketTitle/ticketLink.
bounded-head pageTitle — narrowed out 2026-08-03, read-size work follows /todo/TODO/TODO-002.

Blockers and bummers

A cross-run title cache (path → size/mtime/header line) would cut the remaining 757 opens to ~changed-only; parked pending discussion — do NOT build it under this ticket.

Outcome

landed ///be/#acc80350: todo.js listTopic readdir-derived pages, work.js per-run memos, test/todo/pathreuse repro; board stats 876→13.