Beagle SCM

SUBS-037: JS (jab) submodule recursion for get + post

The JS reference impl (be/, run by jab) recursed submodules for the READ views only (status/diff/log/spot, JAB-004); the WRITE verbs did not. jab get recorded a gitlink then skipped it ("mount recursion is a follow-up", get.js), and jab post had no sub recursion at all. So a nested-submodule clone/edit/ commit cycle diverged from native be. This ticket ports the get mount-recursion and the post recurse+gitlink-bump so jab matches be byte-for-byte across the full cycle, in plain AND --color. Driven by the new cycle gate.

Input

Context

A full get→put/delete→post→get cycle over a 3-level nested submodule repo (aa→bb→cc, path==project) was run with both executors and compared.

• jab get be://…?/aa cloned only the top shard; subs unmounted (get.js:331

  if (kind === "s") return), so bb/bb/cc were missing from the worktree.

• jab post committed the top repo only — no post post:bb/cc recursion, no

  mod bb/cc / mod bb gitlink bump; commit shas diverged from native.

Goals

jab matches native be byte-for-byte on the get/put/delete/post/get cycle.

• get: fetch each sub shard from the SAME remote (project swapped), drop the

  secondary-wt anchor, check the pin tree out, recurse; emit get <sub>?<pin>.

• post: depth-first post-order recursion — post dirty subs first (path-prefixed

  banners, [seg] message tags), bump the parent gitlink (mod <sub>).

• Identical worktrees AND identical output dumps (SGR included) in both modes.

Constraints

• NO C changes — the port lives entirely in be/. Reuse existing kernels

  (wire/ingest/checkout/store, core/recurse.js, fold-decide gitlink bump).

• Clone via be:// (full-store clone, own object store), not a file: wt.
• Commit shas are time-derived; the test ts-pins each level (postcase.sh trick).

WIP

Design decisions

• get: a gitlink leaf calls mountSub (fetch shard → ingest.clone sibling

  shard → write .be anchor → checkout.apply → recurse); sub rows are banner:true so --color full-line-bands them like the seed banner.

• post: handle→postTree (recurse mounted subs in .gitmodules order,

  post-order)→postOne (the old single-repo body). An advanced sub's new tip is injected as a gitlink-bump add decision (injectBumps) so the existing fold-decide/commit path records mod <sub>. fold-commit skips the blob feed for a mode-160000 add (the sha is a commit in the sub shard).

• emit: a generic banner row tag → bannerColumnar in color (also fixes the

  delete: header band); out.color lets post lay 2 blank separators in plain but 1 in color (native's relay quirk).

TODOs

• v get mount-recursion (get.js mountSub + sortGetRows _sub preorder)
• v post recurse + gitlink bump (post.js postTree/postOne/injectBumps)
• v color parity (emit.js banner tag + color export; delete.js header)
• v full-cycle parity test cycle, plain + color, 80/81 green
• POSTNONE sub (clean sub under a dirty parent) emits the post <sub>

  notice — caught+skipped now; the per-level (not single-top) post layout

• local file: parent sub-mount (GET-032 class) — remote path only here

Blockers and bummers

• Gitlink path MUST equal the sub project/shard title for the clean clone

  (else native itself prints HOMENOPROJ fallback noise); the test fixes this.

• be post does NOT push; the cycle's two gets are independent clones (a

  re-get of an ahead wt is SNIFFNOFF), and parity is be-run vs jab-run.

• Pre-existing: be-js-bro-universal is a stale WILL_FAIL xfail that now

  passes (unrelated to this ticket — bro is untouched).

Outcome

LOST — NEVER COMMITTED (confirmed 2026-06-29). The work sat in an uncommitted scratch tree at the commit gate (be/verbs/get/get.js, be/verbs/post/post.js, be/verbs/post/fold-commit.js, be/core/emit.js, be/verbs/delete/delete.js, new be/test/sub/{lib/subcase.sh,cycle/run.sh}, registered be-js-sub-cycle; suite 80/81) but was never landed and that tree was deleted: mountSub now exists in NO tree on disk, no $HOME/todo/SUBS-037 worktree remains, and nothing is in the beagle-ext store. The get+post sub-recursion must be REDONE — owned as net-new work by DIS-058 (D1-D9). REOPEN.

CLOSED — SUPERSEDED (audit 2026-07-07): the redo LANDED under DIS-058 D1-D9 (da0c12c3 + da7fa5a7, be trunk 29Jun): get mounts+recurses subs (shared/submount.js + get.js recurseSubMounts), post recurses post-order + auto gitlink-bump (post.js postTree/postSubs/postOne). The full-cycle gate lives at be/test/sub/cycle (be-js-sub-cycle), green jab-only per d80bdb5c (TEST-003 "js tests pass, no native use"). Remaining sub gaps have their own tickets (SUBS-039/041/045 etc.); nothing left under this code.