Beagle SCM

PUT-007 MED: JS multi-path be put is O(N²) — re-preps + rewrites the whole wtlog per arg

The JS (jab) be put path... handler (be/verbs/put/put.js) processed each path arg independently: for EVERY arg it re-ran stage.prep (a whole working-tree wtScan + the baseline tree read twice) AND commitOps→appendAndAssign (read the whole wtlog, then rewrite the whole wtlog). So be put todo/*/*.mkd (517 matches) "starts anew for each file" — O(N²) in the wtlog rewrite plus O(N×tree) rescans. The JS-side perf analogue of PUT-003's native "per-path, not aggregated".

Input

Context

• Reported: jab put todo/*/*.mkd runs far too long, as if it restarts per file.
• putRun (put.js) looped for (uri of pathUris) putOne(...); putOne did a

  fresh stage.prep(repo, wtlog.open(repo), k) + commitOps PER arg.

• stage.prep scans the ENTIRE wt (classify.wtScan: recursive readdir + lstat)

  and reads the baseline tree twice — none of it arg-specific, redone N times.

• appendAndAssign reads the whole wtlog (ulog.each) then rewrites it whole

  (ulog.write → temp + rename) for EACH arg → O(N²) log rewrites.

• Measured (scratch, native-seeded, all-dirty): 120 files 3.57s, 240 files 12.38s

  (3.5× time for 2× files → super-linear/quadratic).

Goals

• Aggregate a multi-path put into ONE staging pass: prep the engine ONCE, write

  the wtlog ONCE for the whole arg batch; output + rows byte-identical.

• Restore ~linear scaling so a many-file be put completes in well under a second.

Constraints

• No product-behaviour change: same put: rows/skips, same wtlog rows, the

  mtime==row-ts restamp invariant intact; keep the be/ put suite green (Issues).

• Sub-crossing args (SUBS-039) still stage inside the sub's OWN wtlog — they

  can't share the parent engine/commit; leave them per-arg.

WIP

Design decisions

• Hoist stage.prep + wtlog.open above the pathUris loop; reuse one engine

  for every file/dir/move arg (native runs one PUTStage snapshot over all args).

• putOne no longer commits: it classifies against the shared engine and RETURNS

  its ops; putRun concatenates them and calls commitOps ONCE for the batch.

• appendAndAssign left as-is (now called once) — the batching, not an in-place

  append, removes the O(N²); minimal-diff, lower risk.

TODOs

• Repro + baseline timing (120/240 files) proving the quadratic curve.
• Hoist prep; defer per-arg commit; single batched commitOps in putRun.
• Verify be/ put suite + sub/nav/wire-adjacent cases green; re-time.

Blockers and bummers

• ctest can't drive the suite as-is: beagle/build/bin/be (the harness SEED +

  BIN) vanished mid-session; ran cases directly with the build-trace native seeder + build/bin/jab (sanctioned seed, TEST-003 secondary axis).

Outcome

• LANDED 1b7c9934 (be/, 2026-07-06; the commit's own message — "Thanks to

  PUT-007, this was staged and posted by jab" — was staged BY the fix; wtlog: put verbs/put/put.js → post ?#1b7c9934): be/verbs/put/put.js — putRun preps the staging engine once and batches all args into one commitOps; putOne takes the shared engine and returns ops. After: 120 files 0.28s (12.7×), 240 files 0.30s (41×), flat across N (O(N²)→O(N)); correct row counts. Green: put/{bare,file,dir,move,unchanged,dir-clean-unstamped,ref-create, ref-set,ref-set-sha}, sub/{bareput,bareput-nested}, uri011/put-nav-bind, put/wire-refguard.