Upstream freshness: row-share and host-admission observations

Observed 2026-10-23T00:21:01Z. This is a read-only upstream comparison; it does not rebase, fetch into, build, or otherwise mutate the frozen Flare checkout.

Pins and method

Item	Value
Local frozen base	47581ccac82660953a5b59cb04b2750d64edc97d
Origin	https://github.com/NVIDIA/flare-local-inference.git
Upstream default branch	main
Upstream main at observation	e5bd6a67c0ae1d07f4f5dd52b8197cc36f462982
Relationship reported by the official compare API	base is 132 commits behind main and zero commits ahead

The remote ref came from read-only git ls-remote --symref origin HEAD; the symbol comparison used the official main source, because the remote object is intentionally not present in this frozen checkout. Sources: current commit, official comparison, and the repository.

Exact affected surfaces

Candidate surface	Current upstream finding	Contribution consequence
SlabPool::sized_counts row share	slab.rs still calls sized_counts(layout, bytes). The implementation still fixes rows at one quarter of the byte budget (bytes / 4 / row_record) before minimum/budget clamps (lines 334-355). It has no row-share parameter or equivalent configuration knob.	The local row_percent sizing/configuration change is not duplicated upstream.
PrefixCacheOffload configuration	Current PrefixCacheOffload exposes only memory bytes, disk, disk limit, and transport.	No upstream equivalent of the local row_percent or transfer-verification fields was found.
Host-admission observation	Current admit_with_host_store receives scheduler, store, spec, and prefill rate only. Its call site likewise has no metrics argument (lines 1132-1138). The current metrics source has host_cache_fetched_total, but no host_cache_admission_* family or the candidate row/pressure exports.	The raw-match, cost-rejected, capacity-rejected, and fetch-admitted observations are not already upstream under these names or an equivalent producer path.

The present upstream main does retain host_cache_fetched_total; that is the completed physical-fetch statistic already traced in the frozen evidence, not an admission-observation replacement.

Relevant upstream movement

No post-base commit changed slab.rs: its most recent upstream touch remains e0d731e, before the frozen base. Upstream did advance adjacent host-store contracts after the base:

• adc5654, Make multi-host prefix records safe when a peer fails, changes host-store failure handling. In current PrefixStore, poll returns Result<()> and a reporting failure stops the worker so transfer-held ids/pins are not reused. This is not a row-share or admission-counter implementation, but it is an allocator/restore contract change that a future upstream transplant must review.
• c2ee3b7, Forward prefix cache settings to every Orbit rank, affects the CLI forwarding seam. It does not add row share; a future contribution must preserve its current forwarding behavior.
• 01a4e4d, Average planned verification depth across request rounds, changes unrelated metrics.
• The current head, e5bd6a6, starts peer record transport before join and adds multi-host-cache qualification; it does not add an equivalent row-share or admission-observation surface.

The official open issue/PR lists and a focused repository search for row, slab, prefix cache, offload, and admission found no published issue or PR equivalent to the candidate knob/counters. The closed multi-host disk prefix caching PR is related host-cache work, not an equivalent implementation.

Completeness and packaging decision

This comparison is complete for the named sized_counts, PrefixCacheOffload, host-admission producer, and metrics-export surfaces at the observed main SHA. It is partial for all 132 intervening commits and all historical issue/PR discussion: GitHub's comparison file list is capped, and no local remote object was fetched. The focused issue/PR query is a freshness check, not a claim of exhaustive semantic search.

There is no upstream duplicate to remove from a future contribution package. The frozen campaign remains correctly pinned to 47581cca; do not rebase it for this evidence run. If contribution packaging is later authorized, start from then-current main, isolate the row-share/observation delta, and reconcile it with the changed PrefixStore::poll failure contract and current CLI forwarding before proposing anything. No upstream submission was made.