fix(mcp): surface non-ErrNoRows GetNode errors in MemoryWrite

[radimsem]

Closes #218.

Problem

HandleWrite discarded the GetNode error with existing, _ := d.Store.GetNode(ctx, nodeID). Any non-sql.ErrNoRows failure (DB locked, ctx cancelled, corrupted page, broken connection) was silently swallowed, and the handler proceeded to construct a brand-new node from payload defaults. When the operator explicitly supplied an Anchor, that turns a transient store failure into a silent defaulted-metadata create instead of a loud, recoverable error.

Fix

Surface non-ErrNoRows errors while still falling through to the create branch on sql.ErrNoRows, matching the existing pattern in forget.go and summarize.go:

existing, err := d.Store.GetNode(ctx, nodeID)
if err != nil && !errors.Is(err, sql.ErrNoRows) {
    return nil, nil, fmt.Errorf("failed to fetch: node %s: %w", nodeID, err)
}

The early return is before the sole emitNodeChange site, so the one-snapshot-per-call invariant holds and no snapshot is created on the error path.

Test

Adds TestHandleWrite_SurfacesGetNodeError: a cancelled context forces GetNode to return context.Canceled (non-ErrNoRows). Verified regression guard — passes with the fix, fails without (the swallowed path surfaces a later, unrelated emit failure: "failed to write: failed to fetch: head snapshot id").

Verification

• make build clean · go vet ./pkg/mcp/... exit 0 · make test green (220 mcp tests) · make lint 0 issues
• Reviewed by mcp-surface, go-style, and Codex — 0 actionable findings
• Semantics-only fix: tool name/signature/return/locking unchanged, no public-skill update needed

🤖 Generated with Claude Code