Beagle SCM

PACK-004: repack ingest REFUSES a REF_DELTA that arrives in the source pack — thin packs and every non-ofs-delta peer abort

22bec7d6 restored REF_DELTA on the READ side only. PACKResolve/PIDXScanRef take a pack_ref_find so a rotated log's cross-log bases resolve, and REPACK.c:314 EMITS REF_DELTA whenever a crossing delta lands in a new log — but REPACK.c:308 still carries test(obj.type != PACK_OBJ_REF_DELTA, REPACKBASE), so a source pack that CONTAINS a REF_DELTA record aborts with "pack: a delta cites a base that never arrived". That is not an exotic input: a THIN pack names its bases by sha by construction (KEEP-006 §Fetch-side rules says so outright), and so does any peer that did not negotiate ofs-delta. Every incremental fetch is a thin pack, so this bites the moment JAB-020 routes jab get through git.pack. Found by the TEST-004 treadmill; method Issues.

Input

Context

• jab/dog/git/REPACK.c:308 — the else arm of the OFS branch is test(obj.type != PACK_OBJ_REF_DELTA, REPACKBASE): a REF_DELTA record on input ends the run.
• jab/dog/git/REPACK.c:314 — the same loop emits PACK_OBJ_REF_DELTA when baserow->slot != rp->slot, so the writer produces exactly the record shape the reader refuses.
• REPACKRun(fd, buf, shard, conf, ibuf, st) (REPACK.h:72) has no finder parameter; wiring one in is a signature change nothing has made yet.
• Verified 2026-07-27 with TEST-004 test/mill/pack.js: git.git repacked 417,182/417,182 objects clean at a 128 MB cap — which under line 308 PROVES the input pack was OFS-only, not that REF input works.
• KEEP-006 measured that git's own repack emits all-OFS (315,878/315,878 type-6 in git.git), which is why a locally-cloned pack never trips this and a fetched one will.
• The bases a thin pack cites are the ones the client ALREADY HAS — they live in the shard's existing logs, not in the incoming stream, so resolving them needs the shard index too, not just the ingest-local map.

Goals

git.pack ingests a pack containing REF_DELTA records, resolving each base out of the shard rather than aborting.

• REPACKRun takes a base finder (sha → pack slice + offset), the same shape PACKResolve already accepts from 22bec7d6.
• The finder answers from BOTH the ingest-local index being built AND the shard's pre-existing logs, since a thin pack's bases predate the fetch.
• A base that genuinely cannot be found stays a plain-words failure — "a delta cites a base that never arrived" is the right message, just not for every REF record.
• The git.pack binding grows no new argument: the shard path it already takes is what the finder searches.
• TEST-004 pack.js --thin flips from pinning today's refusal to asserting the objects land.

Constraints

• Read-side machinery is already there — reuse pack_ref_find/PIDXScanRef, do not write a second resolver.
• No waiter forest: KEEP-006 established bases always land earlier, and for a thin pack they landed in an earlier FETCH; a lookup suffices.
• ZINFMORE must keep propagating (JAB-020) — do not let the new branch collapse truncation into a base error.
• Errors cross the JS boundary in plain words, never as an ok64 code.

WIP

Design decisions

• The gap is INGEST-only; 22bec7d6's read-side work stands and needs no revisiting.
• Deciding whether the finder is passed in by JS or opened from the shard path inside libdog is open — libdog knows the shard already, and KEEP-006 rules that JS keeps only routing.

TODOs

• Add the finder to REPACKRun's signature and thread it to the delta branch.
• Replace the REPACK.c:308 refusal with a REF resolve: find the base by sha, apply the delta, emit as the base's type.
• Make the finder search the shard's existing logs as well as the run's own index (a thin pack's bases predate the fetch).
• Keep the genuine miss a plain-words error, distinct from a truncated stream.
• TEST-004 pack.js --thin: build a thin pack with git pack-objects --thin and assert the objects land.
• Verify end-to-end through a real incremental jab get, not a unit probe.

Blockers and bummers

• Depends on nothing, but JAB-020 should not declare the clone path converted until this lands: the first clone would work and every incremental fetch would fail.

Outcome

Not started — filed 2026-07-27 from a TEST-004 treadmill finding, after gritzko's read of 22bec7d6.