Struct CapabilityIndexInner 

pub struct CapabilityIndexInner { /* private fields */ }

Secondary index maintained alongside the primary (class, node) → CapabilityMembership store. Three inverted-index dimensions — by tag, by region, by state — matching the plan’s CapabilityIndexInner shape. Powers the fast path for the most common query shapes (find-by-tag, find-in-region, find-by-state) without scanning the full store. Composite queries pick the most selective indexed dimension and filter the others in-memory.

PERF_AUDIT §4.6 — the inner HashSet<(u64, NodeId)> candidate sets use BuildU64TupleHasher (private to this module), a fast multiplicative mixer for (u64, u64) keys. Pre-fix these used the std SipHash default, which paid ~15-25 ns of mixing per insert/contains/remove on keys that are already xxh3-hashed identity bytes (so collision resistance is already there at construction; SipHash’s DoS resistance adds zero protection). The outer HashMap<String, _> / HashMap<NodeState, _> keep the default hasher: tag / region strings come from publishers and the SipHash protection is legitimately relevant there.

Trait Implementations

impl Debug for CapabilityIndexInner

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for CapabilityIndexInner

fn default() -> CapabilityIndexInner

Returns the “default value” for a type.

impl FoldIndex<CapabilityFold> for CapabilityIndexInner

fn index_payload_equivalent( old: &CapabilityMembership, new: &CapabilityMembership, ) -> bool

PERF_AUDIT §4.5 — on_insert keys this index on (tags, derived synthetic tags, region, state). If the two payloads agree on every one of those, an on_remove + on_insert against them nets to a no-op on every bucket (derive_synthetic_index_tags is pure over the payload, so identical inputs produce identical synthetic outputs).

The synthetic tags derive from TWO payload fields: the software.model.* / software.tool.* bundles inside tags (covered by the tags equality) AND the gpu:present / gpu:vendor:<v> projection of hardware — so hardware MUST be part of this comparison or a refresh that changes only the GPU shape would leave by_synthetic stale. Comparing the whole HardwareSummary is slightly conservative (a memory_gb/vram_gb-only delta forces a rebuild the index doesn’t strictly need), but the steady-state refresh the audit targets keeps hardware identical, so the win is unaffected and the check stays future-proof against new hardware-derived synthetic tags. Allow-lists / metadata / price_quote / reflex_addr are NOT consulted by this index and may differ freely.

fn on_insert(&mut self, key: &(u64, NodeId), payload: &CapabilityMembership)

Called after an MergeAction::Insert or MergeAction::Replace commits to the primary store. For Replace, the previous payload was already passed to Self::on_remove.

fn on_remove(&mut self, key: &(u64, NodeId), payload: &CapabilityMembership)

Called before an MergeAction::Replace or an super::Fold::evict_node eviction drops the entry from the primary store, with the payload that’s about to be removed.

fn clear(&mut self)

Drop every cached relation. Called by super::Fold::restore before re-populating from a snapshot.