cxo: fix Cache.cleanDown eviction bug and skip re-encoding unchanged sub-trees

[0pcom]

Summary

Two CXO fixes that together drop a public visor's GC CPU share from ~60% to baseline and substantially reduce live heap. Found via pprof on a public visor running the user-feeds + visor-stats publishers; both fixes apply to every CXO node in the network, not just the visor.

1. cxo/skyobject: Cache.cleanDown never actually evicted anything

pkg/cxo/skyobject/cache.go:341,345 — the filter loop's "skip wanted" / "skip filling" comments said skip, but the code returned, aborting the entire ranking on the first non-evictable entry encountered during map iteration. On a busy node there is virtually always something filling or being awaited, so the cache never evicted: putItem invoked cleanDown once the budget was exceeded, cleanDown built a len(c.is)-sized rank slice and immediately discarded it, and putItem then inserted the new item on top of an already-over-budget cache.

The discarded slice was 95.6% of total bytes allocated under cleanDown in the alloc profile; the un-evicted backlog was the dominant share of live heap under putItem. Fix: continue, not return. Also switched the rank slice to []rankItem (value-slice) to drop a per-entry heap allocation.

This is a pure bug fix and benefits every skyobject.Container — every CXO publisher, subscriber, the TPD aggregator, and the standalone cxo daemon.

2. cxo/treestore: cache encoded sub-node hashes across publishes

Publisher.publishRoot re-walked the entire in-memory tree on every dirty batch and ran encoder.Serialize + entry.Sub.SetValue at every level, even when only a single leaf had changed. The package doc already promised content-addressed re-use of unchanged sub-trees ("only the O(depth) chain of objects from the modified leaf up to the root gets re-encoded"); the implementation didn't match.

Adds a per-memNode encode-cache (pubHash + cached) that's populated after each successful Save and invalidated by putAt/deleteAt/pruneAt along the touched path. encodeNode now places the cached hash directly into the parent's TreeEntry for any unchanged sub-node, skipping the recursive walk and the encoder.Serialize/Cache.Set chain entirely. Container.Save's reference walk increments the rc on the cached hash without descending into it (skyobject/unpack.go:156-167), so the underlying CXO objects stay alive across publishes as long as each successive Root references the same hash.

Cache promotion is deferred until after Save returns nil so an aborted publish doesn't leave the cache pointing at hashes that up.Close has rolled back.

Benefits every treestore.Publisher — pkg/visor/init_stats.go, pkg/visor/cxo_user_feeds.go, cmd/apps/skychat/pairing/pair.go.

Test plan

• go test ./pkg/cxo/... (treestore, skyobject, node, data, idxdb, cxds — all green)
• go build ./...
• New TestEncodeCacheInvalidatesOnlyAffectedPath pins the sibling-preservation contract: a Put under one branch must not change pubHashes on unrelated subtrees, and cached must clear only on the touched path.
• Existing TestPublisherUnchangedSubtreeKeepsHash still passes — the doc-comment guarantee is now actively enforced by the cache fast path rather than only emergent from CXO content-addressing.
• Verify on a live public visor: re-run pprof after deploy and confirm GC CPU share drops and Cache.putItem live heap shrinks.