JSQUE-021: resident session exhausts file descriptors — verb load dies mid-session

Now: DONE

Input

Context

2026-07-18 live, a long-running jab work pager session (the day's WORK-002/WORK-003 clicking): a spell died with err: the verb 'get' failed to load from /home/gritzko/src/journal/work/WORK-003/jsrc/verbs/get/get.js: Error: No — gritzko confirms the full underlying error is "No file descriptors available". The resident process leaked fds until require itself could not open a .js file.

The loader message is the WORK-003 plain-words registry fix doing its job; the status line TRUNCATED the cause at "Error: No" — the full text must survive to the user.
Every click/spell in the resident loop opens stores (packs), wtlogs, dirs; nothing visibly closes them per drive — JSQUE-002's pay-once process turns any per-drive open into a leak.
The native side already has known fd leaks the session would amplify: ABC-013 (FSW watches never removed, kqueue leaks an fd per watch), ABC-014 (PACKClose leaks fd + PAGE mapping + index mmap on error paths).
MEASURED (orchestrator pty probe, 2026-07-18, jab work from work/WORK-003, j/k navigation only): 212 fds at launch, then +1 fd per keypress — 90 keys → 290 fds, ALL 80 new fds are re-opens of the LAUNCH tree's .be (/proc/<pid>/fd readlink); the work view re-opens the wtlog on every redraw and never closes it.
Baseline hog: the 212 launch fds are ~~200 `~~/.be/beagle-ext/*.keeper` pack files ALL held open simultaneously by the store reader — the open set scales with store history, so long sessions start near the ulimit before the per-keypress leak even begins.

Goals

A resident session's fd count stays bounded across arbitrarily many spells/clicks; verb loads never die of fd exhaustion.

Red-first repro: drive a resident session (pty or direct loop harness) through N spells, watch /proc/self/fd count — today it grows per drive; after the fix it plateaus.
Find the actual leak site(s): per-drive store/pack/wtlog/dir opens without close, FSW watches, pty/child fds — measure, don't guess.
Error display: the status line must carry the full cause line, not a width-truncated "Error: No".

Constraints

RULING (gritzko, chat 2026-07-18): "why does jk open a file? should just rerender a buffer" — j/k navigation must RERENDER from the in-memory buffer; no file opens, no data-pass re-run per keypress. Data refresh happens on explicit events (launch, spell completion), not on cursor motion.
Native-side fixes belong to ABC-013/ABC-014 — if the leak is wholly native, this ticket wires the JS-side close discipline and links the rest there; don't fork into libabc without need.
Rulings in force: errors in plain words; no unrelated edits; the JSQUE-020 queue-drop rework may collide — check it before restructuring the loop.

WIP

Design decisions

TODOs

fd-growth repro red as a TEST (probe exists: scratchpad fdtrace.py shape) — N keys → +N fds on the launch .be; assert plateau after fix
wtlog: open once per session (or close per redraw) — the per-keypress .be re-open is the active leak
split out 2026-07-19: native release = ABC-020 (real abc.close unmap binding), JS close+reuse = BE-050 (close after drain, authority/navTree/treeAt memos)
keeper packs: NOT fixable JS-side (abc.close only nulls the GC pin, no deterministic fd close; bounding would re-add churn) — native close(fd)-after-mmap proposal filed in ABC-014
un-truncate the status-line cause text