fix: reject invalid graphs in Import instead of panicking later

[Tyagiquamar]

Problem

Graph.Import trusted the encoded payload structurally. Three classes of malformed input decoded successfully and produced a graph that panics later, far from the cause:

1. Dangling neighbor reference -> nil-pointer panic on Search

A node whose neighbor list references a key absent from its layer got a nil *layerNode stored in its neighbor map (encode.go filled pointers with nodes[key], which silently yields nil on a miss). The first Search over the graph panics:

panic: runtime error: invalid memory address or nil pointer dereference
github.com/coder/hnsw.(*layerNode[...]).search(...)  graph.go:135
github.com/coder/hnsw.(*Graph[...]).search(...)      graph.go:477
github.com/coder/hnsw.(*Graph[...]).Search(...)      graph.go:433

Reproduced against current main by importing a hand-crafted payload (one layer, one node, one neighbor key not present) and then searching.

2. Inconsistent vector dimensionality -> panic inside the distance function

Import enforced no dimensionality constraint, although the doc comment carves out dimensionality as the one thing that must match, and assertDims enforces uniformity for Add. A file containing vectors of different lengths imports cleanly; afterwards Search panics inside vek32 depending on which node map iteration picks as entry point:

panic: slices must be of equal length
github.com/viterin/vek/vek32.CosineSimilarity(...)
github.com/coder/hnsw.CosineDistance(...)  distance.go:14
github.com/coder/hnsw.(*layerNode[...]).search(...)  graph.go:101

The same applies to importing a differently-dimensional file into a populated graph.

3. Negative counts -> panic during decoding

Negative layer/node/neighbor/vector-length varints reach make(...) (make([]*layer[K], -1), make([]K, -3), make([]float32, -4)) and panic inside Import instead of returning an error.

Fix

Validate while decoding and return errors, consistent with the existing version / distance-function checks:

• neighbor keys must exist in their layer (node %v has neighbor %v, but it is not present in its layer)
• every vector must match the dimensionality of the first decoded vector, or of the target graph when it is non-empty
• negative counts are rejected before reaching make
• negative string/vector lengths are rejected in binaryRead

Valid Export output never trips any of the new checks: neighbors are always intra-layer by construction, and assertDims keeps dimensions uniform across a graph.

Verification

New tests (encode_test.go), all passing locally on Go 1.26:

• TestGraph_ImportRejectsDanglingNeighbor - failed with the nil-pointer panic above on unpatched main; now Import errors
• TestGraph_ImportRejectsInconsistentDims/WithinFile - panicked in vek32 on unpatched main; now errors
• TestGraph_ImportRejectsInconsistentDims/AgainstGraph - same for import into a populated graph
• TestGraph_ImportRejectsInvalidCounts/{negativeLayers,negativeNodes,negativeNeighbors,negativeVectorLength} - panicked in make before; now error
• Full existing suite (go test ./..., includes TestGraph_ExportImport roundtrip and TestSavedGraph) passes unchanged

Performance (BenchmarkGraph_Import, 100 nodes x 256 dims): allocations are byte-identical before/after (3016 allocs/op, ~491 KiB/op). Timings on my noisy Windows laptop were inconclusive run-to-run, but the change adds only an ok check to a map lookup that already existed plus integer comparisons per node - no new allocations or passes over the data.

[Tyagiquamar]

Hi @ammario, gentle review ping when you have a chance. This is a focused Import validation fix with regression coverage, and the full Go test suite passes. Happy to adjust anything you'd prefer.

[Tyagiquamar]

Hi @suykerbuyk, could you please review when you have a chance? Thanks!