CODE-024: io.mmap per-file read helpers leak mappings — use the pooled-fd read

Now: OPEN
Sev: MED

Deps: none. The sanctioned replacement already exists — wtread readFileBytes (pooled open/readAll/close, landed under CODE-020 57ff502a). This ticket routes the remaining per-file mmap readers through it. Repro concern is the RESIDENT pager loop over a big tree, per classify's own documented rule.

Input

Context

classify.js:159-165 states the rule: io.mmap leaks the mapping (no JS-side unmap), so a wt with >~1000 tracked files exhausts the process's mmap regions mid-sweep; every later map fails and the file reads as a false mod (the 279-file real-repo regression). Reads MUST use pooled fd.
Offenders still mmap per-file/per-dir: ignore.js:157 (readFileText, per .gitignore during a tree scan), diff.js:317 (readWtFile, per changed file), why.js:27 (per file). ingest.js:242 io._mmap per appendRecords.
In the resident loop these accumulate — the exact regression classify.js documents, just on the diff/search/ignore paths instead of status.

Goals

Route per-file/per-dir reads through the pooled fd path (wtread readFileBytes or a shared read helper), so no mapping outlives the call.

Constraints

Behavior-preserving; pack-oriented mmaps with documented bounds (ingest.js:143/410, 31-bit caveat) are a SEPARATE concern — scope this to the unbounded per-wt-file readers. be/ JS only.

WIP

TODOs

repro/reason: >~1000-file diff or spot search exhausts mappings
ignore.js:157 readFileText → pooled fd read
diff.js:317 readWtFile → wtread readFileBytes
why.js:27 file read → same
ingest.js:242 io._mmap appendRecords — ensure unmap/close or document the bound

Blockers and bummers

diff.js:188 notes an mmap-follows-symlink leak already; fold that concern in.

Outcome

Not started. Suggested commit: CODE-024: per-file reads off io.mmap onto pooled fd.