GIT-003: ingest re-encoder — per-object normalize foreign packs into the OFS-only stream, with dedup

Input

Context

Child of GIT-001; depends on GIT-002. KEEPIngestFile (keeper/KEEP.c:2238) and KEEPIngestStream (:2431) append the received object stream verbatim, patch one PACK count, drop a (count, byte_len) bookmark, then UNPKIndex the slice. A verbatim foreign pack carries REF_DELTA / thin bases / arbitrary offsets — incompatible with the OFS-only invariant. Replace the verbatim append with a dependency-ordered per-object re-emit through the GIT-002 writer: foreign packs may use whatever, the stored result is uniform OFS-only. UNPK already walks and resolves every object (OFS forest + REF waiters + Phase-C thin fallback), so the re-emit fuses into that existing pass rather than adding a second one. See PackLog, CLAUDE.

Verbatim append injects foreign deltas into storage: a received pack's REF_DELTA and thin bases would persist, violating OFS-only and the droppable-shard rule.
Offsets are throwaway: the foreign pack's internal offsets don't match our log, so OFS bases must be re-pointed; a thin REF base may live in an already-stored pack (UNPK Phase C, UNPK.c:525).
No dedup today: whole-pack append re-stores objects the shard already has; per-object ingest can skip them (a storage win on overlapping incremental fetches).
Ordering: streaming arrival order can present a delta before its base for REF/thin; emit must be base-before-dependent.

Goals

Replace the verbatim foreign-pack append with a dependency-ordered per-object re-emit through the GIT-002 writer, so the stored result is uniform OFS-only regardless of foreign format.
Fuse the re-emit into UNPK's existing resolve pass (no second parse) and dedup objects the shard already holds.

Constraints

Quarantine foreign formats in UNPK + this funnel only; the native writer/resolver stay OFS-only (GIT-001 design). Keep UNPK's full REF/OFS/thin reader — it is the foreign reader, not deleted.
Keep deflated bytes where possible: a foreign OFS/REF delta re-emits as OFS by rewriting only the header offset against the base's new log offset (reuse GIT-002's append entry); re-deltify only when the base was dropped by dedup.
Map incoming→local: thread an (incoming-offset / base-sha → local-offset) map for the ingest duration off UNPK's resolved forest — no second parse pass; respect CLAUDE §10 (slices, no ptr math).
RECV push-receive (keeper/RECV.c) feeds the same funnel, so pushed packs normalize identically.

WIP

Design decisions

—

TODOs

Drain → resolve (UNPK forest) → topo-emit per object through the GIT-002 writer.

Re-emit objects in dependency order from UNPK's forest: raw objects appended as-is; OFS/REF deltas re-pointed to the base's local offset and written as OFS_DELTA; thin bases resolved via the existing KEEPGet/Phase-C path then OFS-referenced.
Dedup: before appending, probe the index (KEEPHas); skip objects already present, but still record their local offset in the map so dependents re-point correctly.
Remove the verbatim path: drop the whole-pack u8bFeed + count-patch (:2308-2323) and the single PACK bookmark push (:2360); indexing now happens per appended object inside the writer.
Repro-first (CLAUDE §17): extend test/UNPK.c / test/PSTR.c fixtures — ingest (a) a non-thin OFS pack, (b) a thin REF_DELTA pack whose base is already stored, (c) a pack overlapping existing objects; assert the resulting log is OFS-only, dedup'd, and every object round-trips via KEEPGet. A be↔git clone+incremental-fetch round-trip stays green (build WITH_SSH per GIT-001 note).

Blockers and bummers

GIT-002 dependency satisfied (PACKu8sFeedObj landed e59a1c91).

Outcome

Landed fa52117d on dogs trunk: ingest re-encodes foreign packs to OFS-only via a reemit sink fused into UNPK's resolve pass (keep_ingest_reemit → KEEPPackFeed → PACKu8sFeedObj), with KEEPHas dedup; the verbatim whole-pack append + count-patch + foreign PACK bookmark are removed; RECV funnels through the same path unchanged. Took the re-deltify floor (not deflate-passthrough) per the ticket. build-ssh ctest 394/394 (the 66 ssh-localhost clone/fetch/push round-trips RAN, not skipped); a real 3-fetch incremental store inspected → ref_delta=0, every object round-trips via KEEPGetExact.