Beagle SCM

JS-096: port 10 beagle/test/get store-internals cases to the JS jab get (be/test/get)

The be/test/get JS suite has 4 transport cases (be/file/git/traversal) that assert WORKTREE+stdout+be status PARITY (native vs jab) and deliberately EXCLUDE .be (the JS clone writes a byte-different store). 10 native cases instead assert ON the .be store layout itself — flat shards, peer-URI ref rows, anchor-file shape, no escape. Port those 10 as jab-driven STRUCTURAL store-layout assertions, triaging each: a real assertion where jab get owns the path, a documented N/A (+ split ticket) where it only exercises the C keeper/dog layer jab reuses. Method Issues; harness getcase.

Input

Context

• Native source: `beagle/test/get/{21-nosub-local-refs, 25-remote-shard,

  26-cached-no-wire, 26-remote-shard-refs, 39-be-pathless-project, 40-no-ancestor-wtlog-escape, 43-home-bound-no-escape, 44-file-subs-parent-no-getrow, 48-walk-dot-be-quiet, 54-file-project-selector}`.

• Each native case sources beagle/test/lib/case.sh and asserts on .be/

  (shard dirs, refs rows, wtlog anchor bytes) — NOT a parity diff.

• JS get is be/verbs/get/get.js (SEED → fan-out reconcile); store writes go

  through be/shared/{store,ingest,ulog,wire}.js. parseRemote classifies file:/scheme-less/keeper+local as localish; remotes go seedRemote.

• KEY DIVERGENCE already encoded in getcase tree_eq: the JS clone does NOT

  write ..be.idx and lays a different .be; so these ports CANNOT be byte parity on .be — they are STRUCTURAL invariants (shard exists, row carries //host, .be is a FILE, no remotes/ dir) checked against jab's store.

Goals

• For EACH of the 10: a be/test/get/<name>/run.sh that drives jab get and

  asserts the native case's store-layout invariant against jab's .be.

• Where jab get does not own the path (pure C keeper/dog, or a form jab

  doesn't implement yet), DON'T fake it: document N/A in the ticket with the reason, split a follow-up ticket, and either omit or register WILL_FAIL.

• A new be/test/lib/storecase.sh (sibling to getcase): resolve jab+be,

  hermetic scratch + .be firewall, helpers to inspect .be (shard/row/anchor).

• Register the new cases in be/test/CMakeLists.txt (glob already picks up

  get/*/run.sh); xfail entries go in _BE_WILL_FAIL.

Constraints

• Repro-FIRST per case (CLAUDE.md §17): write the JS case, RUN it, observe

  what jab actually writes, THEN decide port-vs-N/A. No guessing.

• No hand-rolled URI parsing in JS — reuse the URI binding (abc.md rule).
• POSIX sh; no bash. Cases stay file-disjoint (own dir) for clean reconcile.
• Do NOT touch beagle/ (native). Work only in the be/ worktree + its test

  sub. jab/be come from the prebuilt ~/journal/beagle/build/bin on PATH.

WIP

Design decisions

• Triage (CONFIRM each repro-first; this is the starting hypothesis, not gospel):

  - 54 file-project-selector — PORT. Directly tests parseRemote ?/proj shard selection + tip; jab get file:<wt>?/sub → only sub.c, sub's tip. - 26-remote-shard-refs — PORT. ssh clone → assert .be/<proj>/refs carries //localhost, no <proj>/<ref> leaf shard, no remotes/ dir. - 48 walk-dot-be-quiet — PORT. clone a tree with a committed .be entry → real.txt out, jab's checkout/walk emits no bad path '.be' spam. - 40 no-ancestor-wtlog-escape — PORT (adapt). jab get from a fresh subdir writes its OWN .be FILE anchor; ancestor .be/wtlog byte-identical. - 25 remote-shard — ADAPT/xfail. Native writes the shard BEFORE the (failing) fetch; seedRemote fetches first → writes nothing on wire-fail. Port the SUCCEEDING-clone layout arm; the pre-fetch-shard arm is a jab gap → split. - 26-cached-no-wire — N/A→IMPL. //host?ref cached form is unimplemented in parseRemote (treats empty-scheme as local). Split a jab cached-form ticket. - 39 be-pathless-project — N/A. keeper://local?/proj empty-path local-exec is a C keeper(WIRECLI) path jab doesn't run; document + split if wanted. - 44 file-subs-parent-no-getrow — N/A here (SUBS). Sub-source candidate ordering lives in subs.js; route to a SUBS ticket, not this one. - 21 nosub-local-refs — N/A. Asserts the native keeper upload-pack SERVER doesn't rewrite refs; jab reuses that binary — nothing JS to assert. - 43 home-bound-no-escape — N/A. dog/HOME.c walk-up ceiling via strace on be status; C dog-layer, not jab get. Out of scope.

• Net expected: ~4 solid ports (54, 26r, 48, 40), 1 partial (25), ~5 documented

  N/A with split tickets. The agent MUST report the actual repro outcome, not this estimate — a "N/A" that turns out portable should be ported.

TODOs

• Set up the isolated be/ worktree (see Blockers; recipe re-check needed —

  the 2026-06-28 one predates the JAB-001/TEST-003 reshuffle).

• Add be/test/lib/storecase.sh (jab+be resolve, hermetic .be firewall,

  .be-inspection helpers); model env wiring on be/test/lib/getcase.sh.

• 54 — port be/test/get/file-project-selector/run.sh.
• 26r — port be/test/get/remote-shard-refs/run.sh (ssh, WITH_SSH-gated skip).
• 48 — port be/test/get/walk-dot-be-quiet/run.sh (blocked red by JS-097).
• 40 — port be/test/get/no-ancestor-escape/run.sh.
• 25 — port the clone-layout arm; xfail/split the pre-fetch-shard arm.
• 26-cached — RETRIAGE: the cached //host?ref form has since been IMPLEMENTED

  (DIS-055 D7; be/test/get/cached exercises it) — port the store-layout assertions against it instead of splitting an impl ticket.

• 39 / 44 / 21 / 43 — repro, confirm N/A, write the rationale here,

  split follow-up tickets (GET/SUBS/JS) where a real jab gap exists.

• Register new cases in be/test/CMakeLists.txt; xfails into _BE_WILL_FAIL

  (be/test/CMakeLists.txt:29; currently EMPTY — the glob auto-adds get/*/run.sh cases).

• Run the new cases green with prebuilt jab; STOP at the commit boundary,

  report root-cause/triage table/test results for the human gate (no commit).

Blockers and bummers

• Worktree recipe (VERIFIED 2026-06-28; the work ?/be recipe is STALE — the

  shard is beagle-ext, and a pre-made empty .be triggers SNIFFFAIL): mkdir $HOME/todo/JS-096 && cd $HOME/todo/JS-096 && PATH=$HOME/journal/beagle/build/bin:$PATH be get "file:$HOME/.be/?/beagle-ext" — fresh dir, NO pre-made .be; checks out all of be/ (62 ok) and recurses the test sub (mounts test/.be). Never pipe be get to head (SIGPIPE truncates the checkout).

• ssh-transport cases (26r, and 26-cached if pursued) need passwordless ssh to

  localhost; SKIP (exit 0) when absent — mirror getcase.sh's SKIP guards.

Outcome

• (open) Filed 2026-06-28 off the GET test-port plan; worker dispatched into an

  isolated beagle-ext worktree at $HOME/todo/JS-096 — NO port has landed.

• Audit 2026-07-07: be/test/get/ still carries none of the 10 ports and no

  storecase.sh (cases there: be/cached/detach/dirty/file/git/http/oddname/ pin/restore/rewind/traversal); all 10 native source cases still exist under beagle/test/get/. One triage line went stale: the cached //host?ref form landed with DIS-055 (see the retriaged TODO). JS-097 (the 48-port blocker) is also still open.